CA2003337A1 - High-performance computer system with fault-tolerant capability - Google Patents

Abstract

ABSTRACT: A computer system in a fault-tolerant configuration employs multiple identical CPUs executing the same instruction stream, with multiple, identical memory modules in the address space of the CPUs storing duplicates of the same data. The multiple CPUs are loosely synchronized, as by detecting events such as memory references and stalling any CPU ahead of others until all execute the function simultaneously; interrupts can be synchronized by ensuring that all CPUs implement the interrupt at the same point in their instruction stream. Memory references via the separate CPU-to-memory busses are voted at the three separate ports of each of the memory modules. I/O functions are implemented using two identical I/O busses, each of which is separately coupled to only one of the memory modules. A number of I/O processors arc coupled to both I/O busses. I/O devices are accessed through a pair of identical (redundant) I/O processors, but only one is designated to actively control a given device; in case of failure of one I/O processor, however, an I/O device can be accessed by the other one without system shutdown, i.e., by merely redesignating the addresses of the registers of the I/O device under instruction control.

Description

Z00333~

, ; :
:
, .^.
, ...
. .
.... .

.', ~7 ~ 'i ~' '''''~
,:.-,~
. ~:
'.~ ~' .
~.
'`') ' . ~ .

; ''1 RELATED CASES: lbis application discloses subject matter also disclosed in copending U.S. Patent Applications Ser. Nos. 282,469, 282,S38, 282,S40, 283,139 and ~ .

- ~' 283,141, filed December 9, 1988, and Ser. No. 283,S73, filed December 13, 1988, and further discloses subject matter also disclosed in prior copending application Ser. No.
118,503, filed November 9, 1987, all of said applications being assigned to Tandem Computers Incorporated, the assignee of this invention.

BACKGROUND OF THE INVENTION

This invention relates to computer systems, and more particularly to a I/O
processor control in a fault-tolerant multiprocessor system.

Highly reliable digital processing is achieved in various computer architecturesemploying redundancy. For example, TMR (triple modular redundancy) systems may employ three CPUs executing the same instruction stream, along with three separate main memory units and separate I/O devices which duplicate functions, so if one of each type of element fails, the system continues to operate. Another fault-tolerant type of system is shown in U.S. Patent 4,228,496, issued to Katzman et al, for "Multiprocessor Systemn, assigned to Tandem Computers Incorporated. Various methods have been used for synchronizing the units in redundant systems; for ~; example, in said prior application Ser. No. 118,503, filed Nov. 9, 1987, by R. W.
Horst, for "Method and Apparatus for Synchronizing a Plurality of Processors", also assigned to Tandem Computers Incorporated, a method of "loose" synchronizing is ; disclosed, in contrast to other systems which have employed a lock-step synchroniza-tion using a single clock, as shown in U.S. Patent 4,453,215 for "Central Processing Apparatus for Fault-Tolerant Computing", assigned to Stratus Computer, Inc. A
teehnique ealled "synehronization voting" is diselosed by Davies & Wakerly in "Synchronization and Matehing in Redundant Systems", IEEE Transaetions on Com-puters June 1978, pp. 531-539. A method for interrupt synchronization in redundant ` ~ 25 fault-tolerant systems is diselosed by Yondea et al in Proeeeding of 15th Annual Symposium on Fault-Tolerant Computing, lune 1985, pp. 246-251, "Implementation of Interrupt Handler for Loosely Synchronized TMR Systems". U.S. Patent 4,644,498 for "Fault-Tolerant Real Time Clock" diseloses a triple modular redundant clock .,~
; 2 . .~
;' ,. ~ , .

.

., configuration for use in a lMR computer system. U.S. Patent 4,733,353 for "FrameSynchronization of Multiply Redundant Computers" discloses a synchronization method using separately-clocked CPUs which are periodically synchronized by executing asynch frame.

As high-performance microprocessor devices have become available, using higher clock speeds and providing greater capabilities, such as the Intel 80386 and Motorola 68030 chips operating at 25-MHz clock rates, and as other elements of computer systems such as memory, disk drives, and the like have correspondingly become less expensive and of greater capability, the performance and cost of high-reliability processors has been required to follow the same trends. In addition,standardization on a few operating systems in the computer industry in general has vastly increased the availability of applica~ions software, so a similar demand is made ; on the field of high-reliability systems; i.e., a standard operating system must be available.

It is therefore the principal object of this invention to provide an improved high-reliability computer system, particularly of the fault-tolerant type. Another object is to provide an improved redundant, fault-tolerant type of computing system, and one in which high perforrnance and reduced cost are both possible; particularly, it is - preferable that the improved system avoid the performance burdens usually associated with highly redundant systems. A further object is to provide a high-reliabilitycomputer system in which the perforrnance, measured in reliability as well as speed and software compatibiliq, is improved but yet at a cost comparable to other alternatives of lower performance. An additional object is to provide a high-rcliability computer systcm which is capable of executing an operating system which uses virtual memory management with demand paging, and having protected (supervisory or "kernel") mode; particularly an operating system also permittingexecution of multiple processes; all at a high level of performance. Still another object is to provide a high-reliability redundant computer system which is capable of detecting faulty system components and placing them off-line, then reintegrating; 30 repaired system components without shutting down the system.
.' '.
. ~
~: 3 ' .

.. . . . . .
~: ;, ~ . ..
:' . ' , , . ' ' :: , ;
.:

SUMMARY O~ THE rNVENTlON

In accordance with one embodiment of the invention, a computer system employs three identical CPUs typically executing the same instruction stream, and has two identical, self checking memory modules storing duplicates of the same data. A
configuration of three CPUs and two memories is therefore employed, rather than three CPUs and three memories as in the classic TMR systems. Memory re~erences by the three CPUs are made by three separate busses connected to three separate ports of each of the two memory modules. In order to avoid imposing the performance burden of fault-tolerant operation on the CPUs themselves, and imposing the expense, complexity and timing problerns of fault-tolerant clocking, the three CPUs each have their own separate and independent clocks, but are loosely synchronized, as by detecting events such as memory refercnces and stalling any CPU
ahead of others until all execute the function simultaneously; the interrupts are also synchronized to the CPUs ensuring that the CPUs execute the interrupt at the same : 15 point in their instruction stream. The three asynchronous memory references via the separate CPU-to-memory busses are voted at the three separate ports of each of the memory modules at the time of the memory request, but read data is not voted when returned to the CPUs.

` The two memories both perform all write requests received from either the CPUs or the I/O busses, so that both are kept up-to-date, but only one memory module presents read data back to the CPUs or I/Os in response to read requests;the one memory module producing read data is designated the "primary" and the other is the back-up. Accordingly, incoming data is from only one source and is not voted. lhe memoly requests to the two memory modules are implemented while the 2S voting is still going on, so the read data is available to the CPUs a short delay after the last one of the CPUs makes the request. Even write cycles can bc substantially overlappcd bccausc DRAMs uscd for thesc memory modules use a large part of the write access to merely read and refresh, then if not strobed for the last part of the : ' .

.
. .
~' ' .

i ~ .
.

write cycle the read is non-destructive; therefore, a write cycle begins as soon as the first CPU malces a request, but does not complete until the last request has been received and voted ~ood. These featurcs of non-voted read-data rcturns and overlapped accesses allow fault-tolerant operation at high performance, but yet at minimum complexity and expcnse.

1/0 functions are implemented using two identical I/0 busses, each of which is separately coupled to only one of the memory modules. A number of 1/0 processorsare coupled to both I/0 busses, and I/0 devices are coupled to pairs of the 1/0 processors but accessed by only one of the I/0 processors. Since one memory ; 10 module is designated primary, only the I/0 bus for this module will be controlling the I/0 processors, and I/0 traffic between memory module and I/0 is not voted. The CPUs can access the I/0 processors through the rnemory modules (each access being voted just as the memory accesses are voted), but the 1/0 processors can only access the memory modules, not the CPUs; the 1/0 processors can only send interrupts tothe CPUs, and these interrupts are collected in the memory modules before presenting to the CPUs. Thus synchronization overhead for 1/0 device access is not burdening the CPUs, yet fault tolerance is provided. If an 1/0 processor fails, the other one of the pair can take over control of thc 1/0 devices for this 1/0 processor by merely changing the addresses used for the 1/0 device in the 1/0 page table maintained by the operating system. In this manner, fault tolerance and reintegration of an 1/0 device is possible without system shutdown, and yet without hardware expense and performance penalty associated with voting and the like in these I/0paths.

The mcmory system used in the illustrated embodiment is hierarchical at several levels. Each CPU has its own cache, operating at essentially the clock speed of the CPU. Then cach CPU has a local memory not accessible by the other CPUs, and virtual memory management allows the kernel of the operating system and pages .
for the current task to be in local memory for all three CPUs accessible at highspeed without fault-tolerance overhead such as voting or synchronizing imposed. ~ext is the memory module level, referred ~o as global memory, where voting and ~~ 5 `''`
`.'' .

, .. . . . .
~ .

' .

~003337 synchronization take placc 50 some access-time burden is introduced; nevertheless, the speed of the global mcmory is much faster than disk access, so this level is used for page swapping with local memory to keep the most-used data in the fastest area, rather than employing disk for the ffrst level of demand paging.

One of the features of the disclosed embodiment of the invention is ability to replace faulty components, such as CPU modules or memory modules, without shutting down the system. Thus, the system is available for continuous use even though components may fail and have to bc replaced. In addition, the ability to o~tain a high level of fault tolerance with fewcr system components, e.g., no fault-tolerant clocking needed, only two memory modules needed instead of three, voting circuits minimized, etc., means that there are fewer components to fail, and so the reliability is enhanced. That is, there are fewer failures because there are fewer components, and when there are failures the components are isolated to allow thesystem to keep running, while the components can bc replaced without system shut-down.
; ,.
The CPUs of this systcm preferably use a commercially-available high-performance microprocessor chip for which operating systems such as UnixTM are available. The parts of thc system which make it fault-tolerant are either transparent to the operating system or easily adapted to the opcrating system. Accordingly, a ; 20 bigh-performance fault-tolerant system is provided which allows comparability with contemporary widely-usrd mulD-tasking operating system and applirations software.

. . .

;:

~:

;, ~ .
'' :~

BRIEF DESCRI~ION OF THE DR~WINGS

The fcatures believed characteristic of the invention are set forth in the appended claims. The invention itself, however, as well as other features and advantages thereof, may best bc understood by reference to the detailed description of a specific embodiment which follows, when read in conjunction with the accompany-ing drawings, wherein:

Figure 1 is an electrical diagram in block form of a computer system according to one embodiment of the invention;

Figure 2 is an electrical schematic diagram in block form of one of thc CPUs of the system of Figure 1;

Figure 3 is an electrical schematic diagram in block form of one of the microprocessor chip used in the CPU of Figure 2;

Figures 4 and 5 are timing diagrams showing events occurring in the CPU of Figures 2 and 3 as a function of time;

` 15 Figure 6 is an electrical schematic diagram in block form of one of the . mcmo}y modules in the computer system of Figure 1;
.' ~' Figure 7 is a timing diagram shôwing events occwring on the CPU to memo~y busses in the system of Figure 1;

Figure 8 is an ebctrical schematic diagram in block form of one of the I/O
processors in the computer system of Figure 1;

~ Figure 9 is a timing diagram showing events vs. time for the transfer protocol : ~ between a memory module and an l/O processor in the system of Figure 1;
. .

1 `
:, . .

. . ~ .

.~

, ~
. .

Figure 10 is a tirning diagrarn showing events vs. time for execution of irlstructions in the CPUs of Figures 1, 2 and 3;

Figure lOa is a detail view of a part of the diagrarn of Figure 10;

Figures 11 and 12 are tirning diagrams similar to Figure 10 showing events vs.
S time for execution of instructions in the CPUs of Figures 1, 2 and 3;

Figure 13 is an electrical schematic diagram in block form of the interrupt synchronization circuit used in the CPU of Figure 2;

Figures 14, 15, 16 and 17 are tirning diagrams like Figures 10 or 11 showing - events vs. time for execution of instructions in the CPUs of Figures 1, 2 and 3 when an interrupt occurs, illustrating various scenarios;

Figure 18 is a physical memory map of the memories used in the system of Figures 1, 2, 3 and 6;

.. Figure 19 is a virtual memory map of the CPUs used in the system of Figures 1, 2, 3 and 6;

Figure 20 is a diagram of the format of the virtual address and the TLB
entries in the rrueroprocessor chips in the CPU according to Figure 2 or 3;
., Figure 21 is an illustration of the private memory locations in the memory map of the global memory modules in the system of Figures 1, 2, 3 and 6; and ....
; Figure 22 is an electrieal diagram of a fault-tolerant power supply used with . 20 the system of the invention aeeording to one embodiment.
~,,` :' .
:
. .

:~
,; ~ 8 ~ '7 `: '''' .`', . ~
: ~ , ' ..'.-:.,'. , . :
:' ' ' .. ; .

:, . , : :' . : , . : ` : :

DETAILED DESCRIPTION OF SPECIFIC EMBODIMENT

With refcrence to Figure 1, a computer system using features of the invention is shown in one embodiment having three identical processors 11, 12 and 13, referred to as CPU-A, CPU-B and CPU-C, which operate as onc logical processor, all three S typically executing the same instruction stream; the only time the three processors are not exccuting the same instruction stream is in such operations as power-up self test, diagnostics and the like. The three processors are coupled to two memory modules14 and 15, referred to as Memory-#1 and Memory-#2, each memory storing the same data in the same address space. In a preferred embodiment, each one of the processors 11, 12 and 13 contains its own local memory 16, as well, accessible only by the processor containing this memory.

Each one of the processors 11, 12 and 13, as well as each one of the memory modules 14 and 15, has its own separate clock oscillator 17; in this embodiment, the processors are not run in "lock stepn, but instead are loosely synchronized by amethod such as is set forth in the above-mentioned application Ser. No. 118,503, i.e., using events such as external memory references to bring the CPUs into synchroniza-- tion. External interrupts are synchronized among the three CPUs by a technique employing a set of busses 18 for coupling the interrupt requests and status from each of the processors to the other two; each one of the processors CPU-A, CPU-B and CPU-C is responsive to the three interrupt rcquests, its own and the two received from the other CPUs, to present an interrupt to the CPUs at the same point in the ' execution stream. The memory modules 14 and 15 vote the memory references, and allow a memory reference to proceed only when all three CPUs have made the same request (with provision for faults). In this manner, the processors are synchronized at the time of external events (memory references), resulting in the processors typically executing the same instruction stream, in the same sequence, but not necessarilydur~ng aligned clock cycles in the time between synchronization events. In addition, external interrupts are synchronized to be executed at the same point in the instruction stream of each CPU.
;
:, :. 9 ,~ ..
'~

: . .

.
:

~ ' : .. ~ - - .
`, ' ~

The CPU-A processor 11 is connected to the ~.emory-#1 module 14 and to the Memory-#2 module 15 by a bus 21; likewise the CPU-B is connected to the modules 14 and lS by a bus 22, and the CPU-C is connected to the memory modules by a bus 23. These busses 21, 22, 23 each include a 32-bit multiplexed address/data S bus, a command bus, and control lines for address and data strobes. The CPUs have control of these busses 21, æ and 23, so there is no arbitration, or bus-request and bus-grant.

Each one of the memory modules 14 and 15 is separately coupled to a respective input/output bus 24 or 25, and each of these busses is coupled to two (or - 10 more) input/output processors 26 and 27. The system can have multiple I/0 processors as needed to accommodate the I/0 devices needed for the particular system configration. Each one of ~he input/output processors 26 and 27 is connected to a bus 28, which may be of a standard configration such as a VMEbusTM, and each bus 28 is connected to one or more bus irterface modules 29 for interface with a standard I/0 controller 30. Each bus interface module 29 is connected to two of the busses 28, so failure of one I/0 processor 26 or 27, or failure of one of the bus channels 28, can be tolerated. The I/0 processors 26 and 27 can be addressed by the CPUs 11, 12 and 13 through the memory modules 14 and 15, and can signal an interrupt to the CPUs via the memory modules. Disk drives, terminals with CRT
screens and keyboards, and network adapters, are typical peripheral devices operated ` by the controllers 30. The controllers 30 may make DMA-type references to the memory modules 14 and 15 to transfer blocks of data. Each one of the I/0 processors 26, 27, etc., bas certain individual lines directly connected to each one of the memory modules for bus request, bus grant, etc.; tbese point-to-point connections 2S are called "radials" and are included in a group of radial lines 31.

A system status bus 32 is individually connected to each one of the CPUs 11, 12 and 13, to each memory module 14 and 15, and to each of the I/0 processors 26and 27, for the purpose of providing information on the status of each element. This ~i~ status bus provides informadon about which of the CPUs, memory modules and I/0 processors is currently in the system and operadng properly.

.~.

; .
r ~,A
"~
' ' .

~ .
. .~
; ' '~ ' ', ' ' ' ' ' :, ~ :

`

~003337 An acknowlcdge/status bus 33 connecting the thrce CPUs and two memory modules includes individual lines by which the modules 14 and 15 send acknowledge signals to the CPUs when memory requests are madc by tbe CPUs, and at the same time a status field is sent to report on the status of the command and whether it S executed correctly. The memory modules not only check parity on data read from or written to the global memory, but also check parity on data passing through the memory modules to or from the I/O busses 24 and 25, as well as checking the validity of commands. It is through the status lines in bus 33 that these checks are reported to the CPUs 11, 12 and 13, so if errors occur a fault routine can be entered to isolate a faulty component.

Even though both memory modules 14 and 15 are storing the same data in global memory, and operating to perform every memory reference in duplicate, one of these memory modules is designated as primary and the other as back-up, at any given dme. Memory write operations are executed by both memory modules so both are kept current, and also a memory rcad operadon is executed by both, but only the primary module actually loads the read-data back onto the busses 21, 22 and 23, and only the primary memory module controls the arbitration for multi-master busses 24 and 25. To keep the primary and back-up modules executing the same operations, abus 34 conveys control information from primary to back-up. Either module can assume thc role of primary at boot-up, and the roles can switch during operationunder software control; the roles can also switch when selected error conditions are detected by the CPUs or other error-responsive parts of the system.
.
Certain interrupts generated in the CPUs are also voted by the memory modules 14 and 15. When the CPUs encounter such an interrupt condition (and are not staUed), they signal an interrupt request to the memory modules by individual lines in an interrupt bus 35, so the three interrupt requests from the three CPUs can be voted. When aU interrupts have been voted, the memory modules each send a voted-interrupt signal to the three CPUs via bus 35. This voting of interrupts also functions to check on the operadon of the CPUs. The three CPUs synch the voted . ~ .
" 11 :, ., .
"
:

interrupt CPU interrupt signal via the inter-CPU bus 18 and present the interrupt to the processors at a corrunon point in thc instruction stream. This interrupt synchronization is accomplished without stalling any of the CPUs.

CPU Module:

S Referring now to Figure 2, one of the processors 11, 12 or 13 is shown in more detail. All three CPU modules are of the same construction in a preferred embodiment, so only CPU-A will be described here. In order to keep costs within a competitive range, and to provide ready access to already-developed software andoperating systems, it is preferred to use a cornmercially-availablc microprocessor chip, and any one of a number of devices may be chosen. The RISC (reduced instruction set) architecture has some advantage in implementing the loose synchronization as will be described, but more-conventional CISC (complex instruction set) microprocessors such as Motorola 68030 devices or Intel 80386 devices (available in 20-MHz and 25-MHz speeds) could be used. High-speed 32-bit RISC microprocessor devices are available from several sources in three basic types; Motorola produces a device as :, part number 88000, MIPS Computer Systems, Inc. and others produce a chip set referred to as the MIPS type, and Sun Microsystems has announced a so-called SPARCIM type (scalable processor architecture). Cypress Semiconductor of San Jose, California, for example, manufactures a microprocessor referred to as part number ~ 20 CY7C601 providing 20-MIPS (million instructions per second), clocked at 33-MHz, ;~ supporting the SPARC standard, and Fujitsu manufactures a CMOS RISC
;; misroprocessor, part number S-25, also supporting the SPARC standard.

The CPU board or module in the illustrative embodiment, used as an example, employs a microprocessor chip 40 which is in this case an R2000 device designed by MIPS Computer Systems, Inc., and also manufactured by Integrated Device Technology, Inc The R2000 device is a 32-bit processor using RISC architecture to . provide high performance, e.g., 12-MIPS at 16.67-MHz dock rate. Higher-speed versions of this device may be used instead, such as the R3000 that provides 20-MIPS
at 25-MHz clock rate. The processor 40 also has a co-processor used for memory . . ~ .
'`' .

~x ~
. :::' ~i~

,., ' -` ' ' ' .:, ' ' '.... ' ~ .
,.. , ,'-' .. , .. ' ~ ~ ;
.. ~, . . ~ ,-. ... .. .

, . .~. .

~" t;' , , , '' , ' '~, ' ~ , ' ' ': " ' management, including a translation lookaside buffcr to cache translations of logical to physical addresses. The proccssor 40 is coupled to a local bus having a data bus 41, an address bus 42 and a control bus 43. Separate instruction and data cache memories 44 and 45 are coupled to this local bus. These caches are each of 64K-S byte size, for example, and are accessed within a single clock cycle of the processor 40. A numeric or noating point co-processor 46 is coupled to the local bus if additional performance is needed for these types of calculations; this numeric processor device is also commercially available from MIPS Computer Systems as part number R2010. The local bus 41, 42, 43, is coupled to an internal bus structure through a write buffer 50 and a read buffer 51. The write buffer is a commercially available device, part number R2020, and functions to allow the processor 40 to continue to execute Run cycles after storing data and address in the write buffer 50 ; for a write operation, rather than having to execute stall cydes while the write is completing.

In addition to the path through the write buffer 50, a path is provided to allowthe processor 40 to execute write operations bypassing the write buffer 50. This path is a write buffer bypass 52 allows the processor, under software selcction, to perforrn synchronous writes. If the write buffer bypass 52 is enabled (write buffer 50 not - enabled) and the processor executes a write then the~ processor will stall until the write completes. In contrast, when writes are executed with the write buffer bypass 52 disabled the processor will not stall because data is written into the write buffer 50 (unless the write buffer is full). If the writc buffer S0 is enabled when the processor ' ~ 40 pcrforms a write operation, the write buffer 50 captures the output data from bus ~i 41 and the address from bus 42, as well as controls from bus 43. The wnte buffer 50 '~ 2S can hold up to four such data address sets while it waits to pass the data on to the , main memory. The write buffer runs synchronously with the dock 17 of the processor chip 40, so the processor-to-buffer transfers are synchronous and at the machine cycle rate of the processor. The write buffer 50 signals the processor if it is full and unable to accept data Read operations by the processor 40 are checked against the addresses contained in the four-deep write buffer S0, so if a read is attempted to one '::

:.' , .................. , :

: . ' - .
. - . . .
~ ................. . .. .
;'''`''' ` . ' ' :
, . - .

of the data words waiting in the write buffer to be written to memory 16 or to global memory, the read is stalled until the write is completed.
-The write and read bufers S0 and 51 are coupled to an internal bus structure having a data bus 53, an address bus 54 and a control bus 55. The local memory 16 S is accessed by this internal bus, and a bus interface 56 coupled to the internal bus is used to access the system bus 21 (or bus æ or 23 for the other CPUs). The separate data and address busses 53 and 54 of the intemal bus (as derived from busses 41 and 42 of the local bus) are converted to a multiplexcd address/data bus 57 in the system bus 21, and the command and control lines are correspondingly converted to command lines 58 and control lines 59 in this extemal bus.
:,.
The bus interface unit 56 also receives the acknowledge/status lines 33 from the memory modules 14 and 15. In these lines 33, separate status lines 33-1 or 33-2 are coupled from each of the modules 14 and 15, so the responses from both memory ` modules can be evaluated upon the event of a transfer (read or write) between CPUs and global memory, as will be explained.
. .. ~. .
. The local memory 16, in one embodiment, comprises about 8-Mbyte of RAM
which can be accessed in about three or four of the machine cycles of processor 40, and this access is synchronous with the clock 17 of this CPU, whereas the memoryaccess time to the modules 14 and 15 is much greater than that to local memory, and this access to the memory modules 14 and lS is asynchronous and subject to the synchronization overhead imposed by waidng for all CPUs to make the request thenvoting. For comparison, access to a typical commercially-available disk memory throup the 1/0 processors 26, 27 and 29 is measured in milliseconds, i.e., ; ~ considerably slower than access to the modules 14 and 15. Thus, there is a hierarchy 2S of memory access by the CPU chip 40, the highest being the instruction and data caches 44 and 45 which will provide a hit ratio of perhaps 95% when using 64-KByte cache size and suitable fill algorithms. The second highest is the local memory 16, and again by employing contemporary virtual memory management algorithms a hit rado of perhaps 95% is obtained for memory references for which a cache miss .

. . ~
:.`..~
.' ' ;; - . . . . , . . , :
, ,,. . ~........... . :
','~ "' ' ' . ~ ' ' ,-. , .; . . . . -, . . ..
' '; , .

~ .

occurs but a hit in local memory 16 is found, in an example where thc size of the local memory is about 8-MByte. The nct result, from the standpoint of the processor chip 40, is that perhaps greater than 99% of memory references (but not I/O
references) will be synchronous and will occur in either the same machine cycle or in S three or four machine cycles.

Thc local memory 16 is accessed from the internal bus by a memory controller 60 which receives the addresses from address bus 54, and the address strobes from the control bus 55, and generates separate row and column addresses, and RAS andCAS controls, for example, if the local memory 16 employs DRAMs with multiplexedaddressing, as is usually the case. Data is written to or read from the local memory via data bus 53. In additio4 several local registers 61, as well as non-volatile memory 62 such as NVRAMs, and high-speed PROMs 63, as may be used by the operating system4 are accessed by the internal bus; some of this part of the memory is used only at power-o4 some is used by the operating system and may be almost continuously within the cache 44, and other may be within the non-cached part of the memory map.

. .-.
~ External interrupts are applied to the processor 40 by one of the pins of the , ::
control bus 43 or 55 from an interrupt circuit 65 in the CPU module of Figure 2.This type of interrupt is voted in the circuit 65, so that before an interrupt is executed by the processor 40 it is determined whether or not all three CPUs are . presented with the interrupt; to this end, the circuit 65 receives interrupt pending inpuS 66 from the other two CPUs 12 and 13, and sends an interrupt pending signal to the other two CPUs via line 67, these lines being part of the bus 18 connecting the three CPUs 11, 12 and 13 together. Also, for voting other types of interrupts, ; 25 specifically CPU-generated interrupS, the circuit 65 can send an interrupt request from this CPU to both of the memory modules 14 and 15 by a line 68 in the bus 35, then receive separate voted interrupt signals from the memory modules via lines 69 and 70; both memory modules will present the external interrupt to be acted upon.
An interrupt generated in some external sourcc such as a keyboard or disk drive on one of the I/O channels 28, for example, will not be presented to the interrupt pin of .

, .. .
~ ~ .

~,~ . , . ' ' .
~'.''.`~' : . -~`". ~ :; . ' ' :
-~' -~'. ., ~ ;' ' ': .

the chip 40 from the circuit 65 until each one of the CPUs 11, 12 and 13 is at the same point in the instruction stream, as will be explained.

Since the processors 40 are clocked by separate clock oscillators 17, there mustbe some mechanism for periodically bringing the processors 40 back into synchro-nization. Even though the clock oscillators 17 are of the same norninal frequency, e.g., 16.67-MHz, and the tolerance for these devices is about 25-ppm (parts per rnillion), the processors can potentially become many cycles out of phase unlessperiodically brought back into synch. Of course, every time an external interrupt occurs the CPUs will be brought into synch in the sense of bein8 interrupted at the same point in their instruction stream (due to the interrupt synch mechanism), but this does not help bring the cycle count into synch. The mechanism of voting memory references in the memory modules 14 and 15 will bring the CPUs into synch(in real timc), as will be explained. However, some conditions result in long periods where no memory reference occurs, and so an additional mechanism is used to introduce stall cycles to bring the processors 40 back into synch. A cycle counter 71 is coupled to the clock 17 and the control pins of the processor 40 via control bus 43 to count machine cydes which are Run cycles (but not Stall cycles). This counter 71 includes a count register having a maximum count value selected to represent the- period during which the maximum allowable drift ~etween CPUs would occur (taking ~0 into account the specified tolerance for the crystal oscillators); when this count register over~ows action is initiated to stall the faster processors until the slower ; processor or processors catch up. This counter 71 is reset whenever a synchronization is done by a memory reference to the memory modules 14 and 15. Also, a rcfresh counter 72 is employed to perform refresh cycles on the local memory 16, as will be 2S explained. In addition, a counter 73 counts machine cycle which are Run cycles but not Stall cycles, like the counter 71 does but this counter 73 is not reset by amemory reference; the counter 73 is used for interrupt synchronization as explained below, and to this end produces the output signals CC-4 and CC-8 to the interrupt synchronization circuit 65.
,:
,. . .

:;

. , '' ' ' ' ~

.:
?

The processor 40 has a RISC instruction set which does not support memory-to-memory instructions, but instead only memory-to-register or register-to-memory instructions (i.e., load or store). It is important to keep frequently-used data and the currently-executing code in local memory. Accordingly, a block-transfer operation is provided by a DMA state machine 74 coupled to the bus interface 56. The processor 40 writes a word to a register in the DMA circuit 74 to function as a command, and writes the starting address and length of the block to registers in this circuit 74. In one embodiment, the microprocessor stalls while the DMA circuit takes over and executes the block transfer, producing the necessary addresses, commands and strobes on the busses 53-55 and 21. The cornmand executed by the processor 40 to initiate this block transfer can be a read from a register in the DMA circuit 74. Since memory management in the Unix operating system relies upon demand paging, these block transfers will most often be pages being moved between global and local memory and I/O traffic. A page is 4-KBytes. Of course, the busses 21, æ and 23 support single-word read and write transfers between CPUs and global memory; theblock transfers referred to are only possible between local and global memory.
.
The Processor:

: Referring now to Figure 3, the R2000 or R3000 type of microprocessor 40 of the example embodiment is shown in more detail. This device includes a main 32-bit . 20 CPU 75 containing thirty-two 32-bit general purpose registers 76, a 32-bit ALU 77, a zero-to-64 bit shifter 78, and a 32-by-32 multiply/divide circuit 79. This CPU also has a program counter 80 along with associated incrementer and adder. These components are coupled to a processor bus structure 81, which is coupled to the local data bus 41 and to an i~struction decoder 82 with associated control logic to execute instructioDs fetched via data bus 41. The 32-bit local address bus 42 is driven by a ;; virtual memory management arrangement including a translation lookaside buffer (TLB) 83 within an on-chip memory-management coprocessor. The TLB 83 contains sixty-four entries to be compared with a virtual address received from the microprocessor block 75 via virtual address bus 84. The low-order 16-bit part 85 of the bus 42 is driven by the low order part of this virtual address bus 84, and the high-;,., ~

' t ~, ~::

. . .
~ `,~ ' - .
' .,'~ `' .. .
.. . .
. .

order part is from the bus 84 if the ~irtual address is used as the physical address, or is the tag entry from the TLB 83 via output 86 if virtual addressing is used and a hit occurs. The control lines 43 of the local bus are connected to pipeline and bus control circuitry 87, driven from the internal bus structure 81 and the control logic 82.

The microprocessor block 75 in the processor 40 is of the RISC type in that most instructions execute in one machine cycle, and the instruction set uses register-to-register and load/store instructions rather than having complex instructions involving memory references along with ALU operations. There are no complex addressing schemes included as part of the instruction set, such as "add the operand whose address is the sum of the contents of register A1 and register A2 to the operandwhose address is found at the main memory location addressed by the contents of register B, and store the result in main memory at the location whose address isfound in register C." Instead, this operation is done in a number of simple register-to-register and load/store instructions: add register A2 to reg~ster A1; load register B1 from memory location whose address is in register B; add register A1 and register B1; store register B1 to memory location addressed by register C. Optimizing compiler techniques are used to maximize the use of the thirty-two rcgisters 76, i.e., assure that most operations will find the operands already in the register set. The ; load instructions actually take longer tban one machine cycle, and to account for this a latency of one instruction is introduced; the data fetched by the load instruction is not used until tho second cycle, and the intervening cycle is used for some other instruction, if possible.

The main CPU 75 is highly pipelined to facilitate the goal of averaging one instruction e~ecution per machine cycle. Referring to Figure 4, a single instruction is executed o~er a period including five machine cycles, where a machine cycle is one clock period or 60-nsec for a 16.67-MHz dock 17. These Sve cycles or pipe stagesare referred ~o as IF (instruction fetch from I-cache 44~, RD (read operands from register set 76), ALU (perforrn tlle required operation in ALU 77), MEM (access D-cache 45 if required), and WB (write back ALU result to register file 76). As seen in Fig~ure 5, tbese five pipe stages are overl pped sr, tbat in a given rmachine qcle, ~:, ' ~ .
:
. . .
~. .
:. ' 1 . .
.. .. .
. .. - - . .. .
' ' '' ~ .
: .
. - ... - ~
~, .

cycle-5 for example, instruction I#5 is in its first or IF pipe stage and instruction I#l is in its last or WB stage, while the otber instructions arc in the intervening pipe stages.

Memory Module:

With reference to Figure 6, one of the memory modules 14 or 15 is shown in detail. Both memory modules are of the same construction in a preferred embodi-ment, so only the Memory#1 module is shown. The memory module includes three ; input/output ports 91, 92 and 93 coupled to the three busses 21, 22 and 23 coming from the CPUs 11, 12 and 13, respectively. Inputs to these ports are latched into registers 94, 9S and 96 each of which has separate sections to store data, address, command and strobes for a write operation, or address, comrnand and strobes for a read operation. The contents of these three registers are voted by a vote circuit 100 having inputs connected to all sections of all three registers. If all three of the CPUs 11, 12 and 13 make the same memory request (same address, same comrnand), as should be the case since the CPUs are typically executing the same instruction stream, then the memory request is allowed to complete; however, as soon as the first . memory request is latched into any one of the three latches 94, 95 or 96, it is passed ~; on immediately to begin the memory access. To this end, the address, data and ; command are applied to an internal bus including data bus 101, address bus 102 and control bus 103. From this internal bus the memory request accesses various resources, depending upon the address, and depending upon the system configuration.

In one embodiment, a large DRAM 104 is accessed by the internal bus, using a memory controller 105 which accepts the address from address bus 102 and memory request and strobes from control bus 103 to generate multiplexed row and column addresses for the DRAM so that data input/output is provided on the data bus 101.
This DRAM 104 is also referred to as global memory, and is of a size of perhaps 32-` ~ MByte in one embodiment. In addidon, the internal bus 101-103 can access control and status registers 106, a quantity of non-volatile RAM 107, and write-protect RAM
~` 108. The memory reference by the CPUs can also bypass the memory in the memory :,.
.` 19 ' `

. . .

--, :
-- ' ,.
, . . .
. . . . ..
.: , .. . .

. ::

module 14 or 15 and access the I/O busses 24 and ZS by a bus interface 109 whichhas inputs connected to the internal bus 101-103. If the memory module is the primary mcmory module, a bus arbitrator 110 in each memory module controls the bus interface 109. If a memory ~nodule is the backup module, the bus 34 controls the bus interface 109.

A memory access to the DRAM 104 is initiated as soon as the first request is latched into one of the latches 94, 95 or 96, but is not allowed to complete unless the vote circuit 100 determines that a plurality of the requests are the same, with provision for faults. The arrival of the first of the three requests causes the access to the DRAM 104 to begin. For a read, the DRAM 104 is addressed, the sense amplifiers are strobed, and the data output is produced at the DRAM outputs, so if the vote is good after the third request is received then thc requested data is ready for immediate transfer back to the CPUs. In this manner, voting is overlapped with DRAM access.

Referring to Figure 7, the busses 21, 22 and 23 apply memory requests to ports 91, 92 and 93 of the memory modules 14 and 15 in the format illustrated. Each of- these busses consists of thirty-two bidirectional multiplexed address/data lines, thirteen unidirectional command lines, and two strobes. The command lines include a fieldwhich specifies the type of bus activity, such as read, write, block transfer, single transfer, I/O read or write, etc. Also, a field functions as a byte enable for the four bytes. The strobes are AS, address strobe, and DS, data strobe. The CPUs 11, 12 and 13 each control their own bus 21, 22 or 23; in this embodiment, thesc are not multi-mastcr busses, there is no contention or arbitration. For a writc, the CPUdrives the address and comrnand onto the bus in one cycle along with the addressstrobe AS (active low), then in a subsequent gcle (possibly the next cycle, but not necessarily) drives the data onto the address/data lines of thc bus at the samc time as a data strobe DS. The addrcss strobe AS from each CPU causes the address and comrnand then appearing at the ports 91, 92 or 93 to be latchcd into the addrcss and com nand sections of the registers 94, 95 and 96, as these strobes appear, thcn the data strobe DS causes the data to be latchcd. When a plurality (two out of three in ~ : `
, `
:' .

,-.-, ~ ' ,. . ' , ; .i .
.. .
.
. , .
: . - .~, - ~ .. .....
-- . .

this embodimem) of the busses 21, 22 and 23 drive the same memory request into the latches 94, 95 and 96, the vote circuit 100 passes on the final command to the - bus 103 and the memory access will be executed; if the command is a write, an acknowledge ACK signal is sent back to each CPU by a line 112 (specifically line112-1 for Memory#1 and line 112-2 for Memory#2) as soon as the write has been executed, and at tlle same time status bits are driven via acknowledge/status bus 33 (specifically lines 33-1 for Memory#1 and lines 33-2 for Memory#2) to each CPU at time T3 of Figure 7. The delay T4 between the last strobe DS (or AS if a read) and the ACK at T3 is variable, depending upon how many cycles out of synch the CPUs are a~ the dme of the memory request, and depending upon the delay in the votingcircuit and the phase of the internal independent clock 17 of the memory module 14 or 15 compared to the CPU clocks 17. If the memory request issued by the CPUs isa read, then the ACK signal on lines 112-1 and 112-2 and the status bits on lines 33-1 and 33-2 will be sent at the same time as the data is driven to the address/data bus, during time T3; this will release the stall in the CPUs and thus synchronize the CPU chips 40 on the same instruction. That is, the fastest CPU will have executed more stall gcles as it waited for the slower ones to catch up, then all three will be released at the same time, although the clocks 17 will probably be out of phase; the first instrucdon executed by all three CPUs when they come out of stall will be the ` 20 same instruction.
. . .
All data being sent from the memory modulc 14 or 15 to the CPUs 11, 12 and 13, whether the data is read data from the DRAM 104 or from the memory locations106-108, or is l/O data from the busses 24 and 25, goes through a register 114. This register is loaded from the internal data bus 101, and an output 115 from this register , 2S is applied to the address/data lines for busses 21, 22 and 23 at ports 91, 92 and 93 at ~*~ time T3. Parity is checked when the data is loaded to this register 114. All data . written to the DRAM 104, and all data on the I/O busses, has parity bits associated with it, but the parity bits are not transferred on busses 21, 22 and 23 to the CPU
modules. Parity errors detected at the read rcgister 114 are reponed to the CPU via the status busses 33-1 and 33-2. Only the memory module 14 or 15 designated as primary will drive the data in its register 114 onto the busses 21, 22 and 23. The ' . ~
, 21 . ~ ~

, .
.

~;
.; -.,.~,. .. ~ ~
~ .: . .;
~, . . .

. -:

memory module designated as back-up or secondary will complete a read operation all the way up to the point of loading the register 114 and checking parity, and will report status on buses 33-1 and 33-2, but no data will be driven to the busses 21, 22 and 23.

A controller 117 in each memory module 14 or 15 operates as a state machine clocked by the clock oscillator 17 for this module and rcceiving the various command lines from bus 103 and busses 21-23, etc., to generate control bits to load registers and busses, generate external control si~nals, and the like. This controller also is connected to the bus 34 between the memory modules 14 and 15 which transfers status and control information between the two. The controLler 117 in the module 14 or 15 currently designated as primary will arbitrate via arbitrator 110 between the I/O
side (interface 109) and the CPU side (ports 91-93) for access to the common bus101-103. This decision made by the controLler 117 in the primary memory module 14 or 15 is communicated to the controller 117 of other memory module by the lines 34, ; - 1S and forces the other memory module to execute the same access.

The controLler 117 in each memory module also introduces refresh cycles for the DRAM 104, based upon a refresh counter 118 receiving pulses from the clock oscillator 17 for this module. The DRAM must receive 512 refresh cycles every 8-msec, so on average there must be a refresh cycle introduced about every 15-microsec. The counter 118 thus produces an overflow signal to the controller 117every 15-microsec., and if an idle condition exists (no CPU access or l/O accessexccuting) a refresh cycle is implemented by a command applied to the bus 103. If an operation is in progress, the refresh is executed when the current operation is finished. For lengthy operadons such as block transfers used in memory paging, several refresh cycles may be backed up and execute in a burst mode aftcr the transfer is completed; to this end, the number of overflows of counter 118 since the Iast refresh cycle are accumulated in a register associated with the counter 118.

Interrupt requests for CPU-generated interrupts are received from each CPU
11, 12 and 13 individually by Lines 68 in the interrupt bus 35; these interrupt requests ... .
æ
., .

.
. .
.. . .
:- , - . . :- . . .
.
..
:~ .

20~333~
are sent to each memory module 14 and 15. These interrupt request lines 68 in bus 35 are applied to an interrupt vote circuit 119 which compares the three requests and produces a voted imerrupt signal on outgoing line 69 of the bus 35. The CPUs each receive a voted interrupt signal on the two lines 69 and 70 (one from each module 14 and 15) via the bus 35. The voted interrupts from each memory module 14 and 15 are ORed and presented to the interrupt synchronizing circuit 65. The CPUs, under software control, decide which interrupts to service. External interrupts, generated in the 1/0 processors or I/O controllers, are also signalled to the CPUs through the memory modules 14 and 15 via lines 69 and 70 in bus 35, and likewise the CPUs only respond to an interrupt from the primary module 14 or 15.

1/0 Processor:

Referring now to Figure 8, one of the I/O processors 26 or 27 is~shown in detail. The I/O processor has two identical ports, one port 121 to the I/O bus 24 and the other port 122 to the I/O bus 25. Each one of the I/O busses 24 and 25 consists of: a 36-bit bidirectional multiplexed address/data bus 123 (containing 32-bits plus 4-bits parity), a bidirectional command bus 124 defining the read, write, block ~;- read, block write, etc., type of operation that is being executed, an address line that ;' designates which location is being addressed, either internal to I/O processor or on busses 28, and the byte mask, and finally eontrol lines 125 including address strobe, data strobe, address acknowledge and data aeknowledge. The radial lines in bus 31 inelude individual lines from eaeh I/O proeessor to eaeh memory module: bus request from 1/0 proeessor to the memoly modules, bus grant from the memory modules to the I/O proeessor, interrupt request lines from 1/0 proeessor to memory module, and a reset line from memory to I/O proeessor. Lines to indieate whieh memory moduleis prima~y are eonneeted to eaeh I/O processor via the system status bus 32. A
eontroller or state maehine 126 in the I/O proeessor of Figure 8 reeeives the eommand, eontrol, status and radial lines and internal data, and eommand lines from the busses 28, and defines the internal operation of the 1/0 proeessor, ineluding !
j:
': 23 ~`~

~.
t ~`' ~: :

. ~ -:" :' ' -.

X00333'7 operation of latches 127 and 128 which receive the contents of busses 24 and 25 and also hold information for transmitting onto the busscs.

Transfer on the busses 24 and 25 from memory module to I/0 processor uses a protocol as shown in Figure 9 with the address and data separately acknowledged.
The arbitrator circuit 110 in the memory module which is designated primary performs the arbitration for ownership of the I/0 busses 24 and 25. When a transfer from CPUs to I/0 is needed, the CPU request is presented to the arbitration logic 110 in the memory module. When the arbiter 110 grants this request the memory modules apply the address and cornmand to busses 123 and 124 (of both busses 24 and 25) at the same time the address strobe is asserted on bus 125 (of both busses 24 and 25) in time T1 of Figure 9; when the controller 126 has caused the address to be latched into latches 127 or 128, the address acknowledge is asserted on bus 125, then the memory modules place the data (via both busses 24 and 25) on the bus 123 anda data strobe on lines 125 in time T2, following which the controller causes the data to be latched into both latches 127 and 128 and a data acknowledge signal is placed upon the lines 125, so upon receipt of the data acknowledge, both of the memory modules release the bus 24, 25 by de-asserting the address strobe signal. The 1/0 processor then deasserts the address acknow1edge signal.

For transfers from I/0 processor to the memory module, when the I/0 processor needs to use the I/0 bus, it asserts a bus request by a line in the radial bus 31, to both busses 24 and 25, then waits for a bus grant signal from an arbitrator circuit 110 in the primary mcmory modulc 14 or 15, the bus grant line also being one of the radials. When the bus grant has been asscrted, the controller 126 then waits until the address strobe and address acknowledge signals on busses 125 are deasserted (i.e., false) meaning the previous transfer is completed. At that time, thc controller 126 causes the address to bc applied from latches 127 and 128 to lines 123 of both busses 24 and 25, the command to be applied to lines 124, and the address strobe to ; be applied to the bus 125 of both busscs 24 and 25. When address acknowledge is received from both busses 24 and 25, these are followed by applying the data to the ` 24 `'.
..

' ' ' ' ' ` ': . . ' , . '' '- ~ -~003337 address/data busses, along with data strobes, and the transfer is completed with a data acknowledge signals from the memory modules to the I/O processor.

The latches 127 and 128 are coupled to an internal bus 129 including an address bus 129a, and data bus 129b and a control bus 129c, which can address intcrnal status and control registers 130 used to set up the comrnands to be executed by the controller state machine 126, to hold the status distributed by the bus 32, etc.
These registers 130 are addressable for read or write from the CPUs in the address space of the CPUs. A bus interface 131 communicates with the VMEbus 28, under control of the controller 126. The bus 28 includes an address bus 28a, a data bus 28b, a control bus 28c, and radials 28d, and all of these lines are comrnunicated through the bus interface modules 29 to the l/O controllers 30; the bus interface module 29 contains a multiplexer 132 to allow only one set of bus lines 28 (from one 1/0 processor or the other but not both) drive the controller 30. Internal to the controller 30 are command, control, status and data registers 133 which (as is standard practice for peripheral cantrollers of this type) are addressable from the CPUs 11, 12 and 13 for read and write to initiate and control operations in I/O
devices.

Each one of the I/O controllers 30 on the VMEbuses 28 has connections via a multiplexer 132 in the BIM 29 to both I/O processors 26 and 27 and can be : 20 controlled by either one, but is bound to one or the other by the program executing in the CPUs. A particular address (or set of addresses) is established for control and data-transfer registers 133 representing each controller 30, and these addresses are maintained in an I/O page table (normally in the kernel data section of local memory) by the operating system. These addresses associate each controller 30 asbeing accessible only through either I/O processor #1 or #2, but not both. That is, a different address is used to reach a particular register 133 via I/O processor 26 compared to I/O processor 27. The bus interface 131 (and controller 126) can switch `d the multiplexer 132 to accept bus 28 from one or the other, and this is done by a write to the registers 130 of the I/O processors from the CPUs. Thus, when the device driver is called up to access this controller 30, the operating system uses these ':
.

, -'. ' : .

addresses in the page table to do it. The processors 40 access the controllers 30 by 1/0 writes to the control and data-transfer registers 133 in these controllers using the write buffer bypass path 52, rather than through the write buffer 50, so these are synchronous writes, voted by circuits 100, passed through the memory modules to the busses 24 or 25, thus to the selected bus 28; the processors 40 stall until the write is completed. The I/0 processor board of Figure 8 is configured to detect certain failures, such as improper cormnands, time-outs where no response is received over VMEbus 28, parity-checked data if implemented, etc., and when one of these failures is detected the I/0 processor quits responding to bus traffic, i.e., quits sending address acknowledge and data acknowledge as discussed above with reference to Figure 9. This is detected by the bus interface 56 as a bus fault, resulting in an inter~upt as will be explained, and self-correcting action if possible.

Error Recovery:

The sequence used by the CPUs 11, 12 and 13 to evaluate responses by the memory modules 14 and 15 to transfers via busses 21, 22 and 23 will now be described. This sequence is defined by the state machine in the bus interface units 56 and in code executed by the CPUs.

In case one, for a read transfer, it is assumed that no data errors are indicated in the status bits on lines 33 from the primary memory. Here, the stall begun by the memory reference is ended by asserting a Ready signal via control bus S5 and 43 to allow instrucdon execudon to continue in each microprocessor 40. But, another transfer is not started until acknowledge is received on line 112 from the other (non-primary) memory module(or it times out). An interrupt is posted if any error wasdetected in either status field (lines 33-1 or 33-2), or if the non-primary memory times out.
, In case two, for a read transfer, it is assumed that a data error is indicated in the status lines 33 from the primary memory or that no response is received &om the ` :`
~ 26 :, i ', .

' . . . :'`' , ~
~ . -,'- ~ ' ; ~

primary memory. The CPUs will wait for an acknowledgc from the other memory, and if no data errors are found in status bits from the other memory, circuitry of the bus interface 56 forces a change in ownership (primary memory status), then a retry is instituted to see if data is correctly read &om the new primary. If good status is S received from the new primary, then the stall is ended as before, and an interrupt is posted to update the system (to note one memory bad and different memory is primary). However, if data error or timeou~ results from this attempt to read from the new primary, then an interrupt is asserted to the processor 40 via control bus 55 and 43.

For write transfers, with the write buffer 50 bypassed, case one is where no data errors are indicated in status bits 33-1 or 33-2 from the either memory module.
The stall is ended to allow instruction execution to continue. Again, an interrupt is posted if any error was detected in - ither status field.

For write transfers, write buffer 50 bypassed, case two is where a data error isindicated in status from the primary memory, or no response is received from theprimary memory. The interface controller of each CPU waits for an acknowledge from the other memory module, and if no data errors are found in the status fromthe other memory an ownership change is forced and an interrupt is posted. But if data errors or timeout occur for the other (new primary) memory module, then an interrupt is asserted to the processor 40.

For write transfers with the writc buffer 50 enabled so the CPU chip is not stalled by a write operation, case one is with no errors indicated in status from either memory module. The transfer is endcd, so another bus transfer can begin. But if any error is detected in either status field an interrupt is posted.

` 25 For write transfers, write buffer 50 enabled, case two is where a data error is indicated in status from the primary memory, or no response is receiwd from the primary memory. The mechanism waits for an acknowledge from the other memory, and if no data error is found in the status from the other memory thcn an ownership .:`
~ 27 ', .~ . .
, , ' ' " . .

change is forced and an interrupt is posted. But if data error or timeout occur for the other mcmory, then an interrupt is posted.

Once it has been deterrnined by the mechanism just described that a memory module 14 or 15 is faulty, the fault condition is signalled to thc operator, but the S system can continue operating. The operator will probably wish to replace the memory board containing the faulty module, which can be done while the system ispowered up and operating. The system is then able to re-integrate the new memoryboard without a shutdown. This mechanism also works to revive a memory module that failed to execute a write due to a soft error but then tested good so it need not be physically replaced. The task is to get the memory module back to a state where its data is identical to the other memory module. This revive mode is a two stepprocess. First, it is assumed that the memory is uninitialized and may contain parity errors, so good data with good parity must be written into all locations, this could be all zeros at this point, but since all writes are executed on both memories the way this first step is accomplished is to read a location in the good memory module then write this data to the sarne location in both memory modules 14 and 15. This is done while ordinary operations are going on, interleaved with the task being performed. Writes originating from the I/O busses 24 or 25 are ignored by this revive routine in its first stage. After all locations have been thus written, the next step is the same as the first except that I/O accesses are also written; that is, I/O
` writes from the I/O busses 24 or 25 are executed as they occur in ordinary traffic in the executing task interleaved with reading every location in the good memory and writing this same data to the same locadon in both memory modules. When the modules have been addressed from zero to maximum address in this second step, the memories are idendcal. During this second revive step, both CPUs and I/O
processors expect the memory module being revived to perform all operations without errors. The I/O processors 26, 27 wiU not use data presented by the memory module being revived during data read transfers. After completing the revive process the revived memory can then be (if necessary) designated primary.

.:

: ' ., , .. . .
i:.

, . ';~ . -. '. : : : : ' . .
: . . . .

' ' r ~ ' ' ' ~003337 A similar revive process is provided for CPU modules. When one CPU is detected faulty (as by the memory voter 100, etc.) the other two continue to operate, and the bad CPU board can be replaced without system shutdown. When the new CPU board has run its power-on self-test routines from on-board ROM 63, it signals S this to the other CPUs, and a revive routine is executed. First, the two good CPUs will copy their state to global memory, then all three CPUs will execute a "soft reset"
whereby the CPUs reset and start executing from their initialization routines in R~M, so they will all come up at the exact same point in their instruction stream and will be synchronized, then the saved state is copied back into all three CPUs and the task previously executing is continued.

As noted above, the vote circuit 100 in each memory module determines whether or not all three CPUs make identical memory references. If so, the memory operation is allowed to proceed to completion. If not, a CPU fault mode is entered.
The CPU which transmits a different memory reference, as detected at the vote circuit 100, is identified in the status returned on bus 33-1 and or 33-2. An interrupt is posted and a software subsequently puts the faulty CPU offline. This offline status ~ is reflected on status bus 32. The memory reference where the fault was detected is ; allowed to complete based upon the two-out-of-three vote, then until the bad CPU
- board has been replaced the vote circuit 100 requires two identical memory requests from the two good CPUs before allowing a memory reference to proceed. The system is ordinarily configured to continue operating with one CPU off-line, but not : two. However, if it werc desired to operate with only one good CPU, this is an altcrnative available. A CPU is voted faulty by the voter circuit 100 if different data is detected in its memory request, and also by a time-out; if two CPUs send identical memory requests, but thc third docs not send any signals for a preselected time-out period, that CPU is assumed to be faulty and is placed off-line as before.
.
The I/O arrangement of the system has a mechanism for software reintegra-tion in the event of a failure. That is, the CPU and memory module core is hardware fault-protected as just described, but the I/O portion of the system issoftware fault-protected. When one of the I/O processors 26 or 27 fails, the ..

`' .. .
, . .

;.

controllers 30 bound ~o that 1/0 processor by softwarc as mentioned above are switched over to the other I/0 processor by software; the operating system rewrites the addresses in the I/0 page table to use the new addresses for the same controllers, and from then on these controUers are bound to the other one of the pair S of I/0 processors 26 or 27. The error or fault can be detected by a bus error terminating a bus cycle at the bus interface 56, producing an exception dispatching into the kernel through an exception handler routine that will determine the cause of the exception, and then (by rewriting addresses in the I/0 table) move all the controllers 30 from the failed I/0 processor 26 or 27 to the other one.

When the bus interface 56 detects a bus error as just described, the fault must be isolated before the reintegration scheme is used. When a CPU does a write, either to one of the I/0 processors 26 or 27 or to one of the 1/0 controllers 30 on one of the busses 28 (e.g., to one of the control or status registers, or data registers, in one of the I/0 elements), this is a bypass operation in the memory modules and both memory modules execute the operation, passing it on to the two I/0 busses 24 and 25; the two I/0 processors 26 and 27 both monitor the busses 24 and 25 and check parity and check the cornmands for proper syntax via the controllers 126. For example, if the CPUs are executing a write to a register in an I/0 processor 26 or 27, if either one of the memory modules presents a valid address, valid cornmand and valid data (as evidenced by no parity errors and proper protocol), the addressed 1/0 processor will write the data to the addressed location and respond to the memory module with an Acknowledge indication that the write was completed successfully.Both memory modules 14 and 15 are monitoring the responses from the I/0 processor 26 or 27 (i.e., the address and data acknowledge signals of Figure 9, and associated status), and both memory modules respond to the CPUs with operation status on lines 33-1 and 33-2. (If this had been a read, only the primary memorymodule would return data, but both would return status.) Now the CPUs can deterrnine if both executed the write correctly, or only one, or none. If only one returns good status, and that was the primary, then there is no need to force an; 30 ownership change, but if the backup returned good and the primary bad, then an ownership change is forced to make the one that executed correctly now the primary.
" ' ;:`
,' ' .~' '' ; ' '- '' , ' ~ . . . . . .
. i . . .
: ' ' ,- ' , , ' . , ~ , -, : .
. . .
: .: -, In either case an interrupt is entered to report the fault. At this point the CPUs do not know whether it is a memory module or something downstream of the memory modules that is bad. So, a si nilar write is attempted to the other I/O processor, but if this succeeds it does not necessarily prove the memory module is bad because the I/O processor initially addressed could be hanging up a line on the bus 24 or 25, for example, and causing parity errors. So, the proces-s can then selectively shut off the I/O processors and retry the operations, to see if both memory modules can correctly execute a write to the same I/O processor. If so, the system can continue operating with the bad I/O processor off-line until replaced and reintegrated. But if the retry still gives bad status from one memory, the memory can be off-line, or further fault-isolation steps taken to make sure the fault is in the memory and not in some other element; this can include switching all the controllers 30 to one I/O processor 26 or 27 then issuing a reset command to the off I/O processor and retry communicationwith the online I/O processor with both memory modules live - then if the reset I/O
processor had been corrupting the bus 24 or 25 its bus drivers will have been turned off by the reset so if the retry of cormnunication to the online I/O processor (via both busses 24 and 25) now returns good status it is known that the reset I/O
processor was at fault. In any event, for each bus error, some type of fault isolation - sequence in implemented to determine which system component needs to be forced offline.

;.,, Synchronization:

The processors 40 used in the illustrative embodiment are of pipelined architecture with overlapped instrucdon execution, as discussed above with reference to Figures 4 and 5. Since a synchronization technique used in this embodiment relies upon cycle counting, i.e., incrementing a counter 71 and a counter 73 of Figure 2 every time an instruction is executed, generally as set forth in application Ser. No.
118,503, there must be a definition of what consdtutes the execution of an instruction in the processor 40. A straightforward definidon is that every time the pipeline`~ advances an instruction is executed. One of the control lines in the control bus 43 is ~ `

~ 31 .. .
~;.
~;:

.. . ...

t:
~ '' , ' . ' .
. . `

. . .

a signal RUN# wh;ch indicates that the pipeline is stalled; when RU~# is high the pipeline is staUed, when RUN# is low (logic zero) the pipeline advances cach machine cycle. ~his RUN# signal is used in the nu neric processor 46 to monitor the pipeline of the processor 40 so this coprocessor 46 can run in lockstep with itsassociated processor 40. This RUN# signal in the control bus 43 along with the clock 17 are used by the counters 71 and 73 to count Run cycles.

The size of the counter register 71, in a preferred embodiment, is chosen to be 4096, i.e., 2l2, which is selected because the tolerances of the crystal oscillators used in the clocks 17 are such that the drift in about 4K Run cycles on average results in a skew or difference in number of cycles run by a processor chip 40 of about all that can be reasonably allowed for proper operation of the interrupt synchronization as explained below. One synchronization mechanism is to force action to cause the CPUs to synchronize whenever the counter 71 overflows. One such action is to force a cache miss in response to an overflow signal OVFI, from the counter 71; this can be done by merely generating a false Miss signal (e.g., TagValid bit not set) oncontrol bus 43 for the next I-cache reference, thus forcing a cache miss exception routine to be entered and the resultant memory reference will produce synchronization just as any memory reference does. Another method of forcing .` synchronization upon overflow of counter 71 is by forcing a stall in the processor 40, which can be done by using the overflow signal OV~;L to generate a CP Busy (coprocessor busy) signal on control bus 43 via logic circuit 71a of Figure 2; this CP
Busy signal always resultt in the procestor 40 entering stall until CP Busy is deatscrted. All three procettort will enter thit stall becaute they are executing the ; same code and will count the same cycles in their counter 71, but the actual time they enter the ttall will vsry; the logic circuit 71a receiws the RUN# signal from bus 43 of the other two procetsort via input R#, so when all three have stalled the CP
Busy signal it releated and the processort will come out of stall in synch again.
.~ .
Thut, two synchronization techniques have been described, the first being the synchronization resulting from voting the memory references in circuits 100 in the memory modules, and the second by the overflow of counter 71 as just set forth. In ~; 32 .,;
. . .
:., `'` ' - .

. :, .
.: , ' ' .
:.~ ' .

addition, interrupts are synchronized, as will be described below. It is important to note, however, that the processors 40 are basically running free at their own clock speed, and are substantially decoupled from one another, except when synchronizing events occur. The fact that microprocessors are used as illustrated in Figures 4 and 5 S would make lock-step synchronization with a single clock more difficult, and would degrade perfommance; also, use of the write buffer 50 serves to decouple the processors, and would be much less effective with close coupling of the processors.
Likewise, the high-performance resulting from using instruction and data caches, and virtual memory management with the TLBs 83, would be more difficult to implementif close coupling were used, and perfomlance would suffer.
;' The interrupt synchronization technique must distinguish between real time and so-called "virtual time". Real time is the extemal actual time, clock-on-the-wall time, measured in seconds, or for convenience, measured in machine cycles which are 6~nsec divisions in the example. The clock generators 17 each produce clock pulses in real time, of course. Virtual time is the intemal cycle-count time of each of the processor chips 40 as measured in each one of the cycle counters 71 and 73, i.e., the instruction number of the instruction being execused by the processor chip, measured in instructions since some arbitrary beginning point. Referring to Figure 10, the relationship between real time, shown as to to tl2, and virtual time, shown as instruction number (modulo-16 count in count register 73) lo to Il5, is illustrated.
Each row of Figure 10 is the cycle count for one of the CPUs A, B or C, and eachcolumn is a "point" in real tirne. The clocks for the CPUs will most likely be out of phase, so the actual time correlation will be as seen in Fiure lOa, where the instruction numbers (colu nns) are not perfectly aligned, i.e., the cycle-count does not 2S change on aligned real-time machine cycle boundaries; however, for explanatory purposes the illustration of Figure 10 will suffice. In Figure 10, at real time t3 the CPU-A is at the third instruction, CPU-B is at count-9 or executing the ninth instruction, and CPU-C is at the founh instruction. Note that both real time andvinual timc can only advancc.

. .
, .
, 33 -;
, . .

' , ;:
, ., Z00333~
The processor chip 40 in a CPU stalls under certain conditions when a resource is not available, such as a D-cache 45 or l-cache 44 miss during a load or an instruction fetch, or a signal that the write buffer 50 is full during a store operation, or a "CP Busy" signal via the control bus 43 that the coprocessor 46 is busy (the coprocessor receives an instruction it cannot yet handle due to data dependency or limited processing resources), or the multiplier/divider 79 is busy (the internal multiply/divide circuit has not completed an operation at the time the processorattempts to access the result register). Of these, the caches 44 and 45 are "passive resources" which do not change state without intervention by the processor 40, but the remainder of the items are active resources that can change state while the processor is not doing anything to act upon the resource. For example, the write buffer 50 can change from full to empty with no action by the processor (so long as the processor does not perform another store operation). So there are two types of stalls: stalls on passive resources and stalls on active resources. Stalls on active resources are called interlock stalls.

Since the code streams executing on the CPUs A, B and C are the same, the states of the passive resources such as caches 44 and 45 in the three CPUs are necessarily the same at every point in virtual timc. If a stall is a result of a conflict at a passive resource (e.g., the data cache 45) then all three processors will perform a ;i 20 stall, and the only variable will be the length of the stall. Referring to Figure 11, assume the cache miss occurs at I~, and that the access to the global memory 14 or lS resulting from the miss takes eight clocks (actually it may be more than eight). In this case, CPU-C begins the access to global memory 14 and 15 at tl, and the controller 117 for global memory begins the memory access when the first processor 2S CPU-C~ signals the beg~nning of the memory access. The controller 117 completes tne access eight clocks later, at t~p although CPU-A and CPU-B each stalled less than the eight clocks required for the memory access. The result is that the CPUs become synchronized in real time as well as in virtual time. This example also illustrates the advantage of overlapping the access to DRAM 1û4 and the voting in circuit 100.
~ x , .
. i ~ :' .
.~
~::

, . :; .
....
.. . .
',,:',: ,. , : ~ .
. ~
.
.
:
; .- . , .

-, ~ .. :.
.

~003337 Interlock stalls present a different situation from passive resource stalls. OneCPU can perform an interlock stall when another CPU does not stall at all.
Referring to Figure 12, an interlock stall caused by the write buffer 50 is illustrated.
The cycle-counts for CPU-A and CPU-B are shown, and the full flags A,~" and B,"bfrom write buffers 50 for CPU-A and CPU-B are shown below the cycle-counts (highor logic one means full, low or logic zero means empty). The CPU checks the state of the full flag every time a store operation is executed; if the full flag is set, the CPU stalls until the full flag is cleared then completes the store operation. The write buffer 50 sets the full flag if the store operation fills the buffer, and clears the full flag whenever a store operation drains one word from the buffer thereby freeing a location for the next CPU store operation. At time t0 the CPU-B is three clocks ahead of CPU-A, and the write buffers are both full. Assume the write buffers are performing a write operation to global memory, so when this write completes during t5 the write buffer full flags will be cleared; this clearing will occur synchronously in t6 in real time (for the reason illustrated by Figure 11) but not synchronously in virtual time. Now, assume the instruction at cycle-count I6 is a store operation; CPU-A
executes this store at t6 after the write buffer full flag is cleared, but CPU-B tries ~o execute this store operation at t3 arld finds the write buffer full flag is still set and so has to stall for three clocks. Thus, CPU-B performs a stall that CPU-A did not.

The property that one CPU may stall and the other not stall imposes a restriction on the interpretation of the cycle counter 71. In Figure 12, assume interrupts are presented to the CPUs on a cycle count of I~ (while the CPU-B is stalling from the I6 instruction). The run cyck for cycle count I, occurs for both CPUs at t~. If the cycle counter alone presents the interrupt to the CPU, then CPU-A would see the interrupt on cycle count I, but CPU-B would see the interrupt ` during a stall cycle resulting from cycle count I6, so this method of presenting interrupts would cause the two CPUs to take an exception on different instructions, a condition that would not have occurred if either all of the CPUs stalled or nonestalled.
'~`, : . .
:.~

"' `:

- - . .
. ~ . .

. . .
..... ~ ~ . . . .
' ' ' , .

Another restriction on the interpretation of the cycle counter is that there should not be any delays betwcen detecting the cycle count and performing an action.
Again rcferring to Figure 12, assume interrupts are presented to the CPUs on cycle count I6, but because of implementation restrictions an extra clock delay is interposed between detection of cycle count I6 and presentation of the interrupt to the CPU.
The result is that CPU-A sees this interrupt on cycle count 1~, but CPU-B will see the interrupt during the stall from cycle count I6, causing the two CPUs to take an exception on different instructions. Again, the importance of monitoring the state of the instruction pipeline in real time is illustrated.

Interrupt Synchronization:

The three CPUs of the system of Figures 1-3 are required to function as a single logical processor, thus requiring that the CPUs adhere to certain restrictions regarding their internal state to ensure that the programming model of the threeCPUs is that of a single logical processor. Except in failure modes and in diagnostic functions, the instruction streams of the three CPUs are required to be identical. If not identical, then voting global memory accesses at voting circuitry 100 of Figure 6 would be difficult; the voter would not know whether one CPU was faulty or whether it was executing a different sequence of instructions. The synchronization scheme is designed so that if the code stream of any CPU diverges from the code stream of the other CPUs, then a failure is assumed to have occurred. Interrupt synchronization ~- provides one of the mechanisms of maintaining a single CPU image.

t All interrupts are required to occur synchronous to virtual time, ensuring that i~' the instruction streams of the three processors CPU-A, CPU-B and CPU-C will not diverge as a result of interrupts (there are other causes of divergent instruction 2S streams, such as one processor reading different data than the data read by the other processors). Several scenarios exist whereby interrupts occurring asynchronous to virtual time would cause the code streams to diverge. For example, an interrupt causing a context switch on one CPU before ptOCCSS A completes, but causing the ~. ~
~.
~ 36 ..~
'~

. .

,~ . .
,.............. .

. . ~
~:

200333'7 context switch after process A completes on another CPU would result in a situation where, at some point later, one CPU continues executing process A, but the otherCPU cannot execute process A because that process had already completed. If in this case the interrupts occurred asynchronous to virtual time, then just the fact that the exception program counters were different could cause problems. The act of writing the exception program counters to global memory would result in the voter detecting different data from the three CPUs, producing a vote fault.

Certain types of exceptions in the CPUs are inherently synchronous to virtual time. One example is a breakpoint exception caused by the execution of a breakpoint instruction. Since the instruction streams of the CPUs are identical, the breakpoint exception occurs at the same point in virtual time on all three of the CPUs.
Similarly, all such internal exceptions inherently occur synchronous to virtual time.
For example, TLB exceptions are internal exceptions that are inherently synchronous.
TLB exceptions occur because the virtual page number does not match any of the entries in the TLB 83. Because the act of translating addresses is solely a function of the instruction stream (exactly as in the case of the breakpoint exception), thetranslation is inherently synchronous to virtual time. In order to ensure that TLB
exceptions are synchronous to virtual time, the state of the TLBs 83 must be identical in all three of the CPUs 11, 12 and 13, and this is guaranteed because the TLB 83 can only be modified by software. Again, since all of the CPUs execute the same instruction stream, the state of the TLBs 83 are always changed synchronous to virtual time. So, as a general rule of thumb, if an action is performed by software then the action is synchronous to virtual time. If an action is performed by hardware, which does not use the cycle counters 71~ then the action is generally synchronous to real time.
.
External exceptions are not inherently synchronous to virtual time. I/O devices 26, 27 or 30 have no information about the virtual time of the three CPUs 11, 12 and `. 13. Therefore, all interrupts that are generated by these I/O devices must be synchronized to virtual time beforc presenting to the CPUs, as explained below.

: ' " .

.' . . . . .
.

, ' .

Floating point exceptions are different from l/O device interrupts because the ûoating point coprocessor 46 is tightly coupled to the microprocessor 40 within the CPU.
External devices view the three CPUs as one logical processor, and have no information about the synchronaity or lack of synchronaity between the CPUs, so the external devices cannot produce interrupts that are synchronous with the individual instruction stream (virtual time) of each CPU. Without any sort of synchronization, if some external device drove an interrupt at tirne real time tl of Figure 10, and the interrupt was presented directly to the CPUs at this time then the three CPUs would take an exception trap at different instructions, resulting in an unacceptable state of the three CPUs. This is an example of an event (assertion of an interrupt) which is synchronous to real time but not synchronous to virtual time.
.. :
Interrupts are synchronized to virtual time in the system of Figures 1-3 by - -performing a distributed vote on the interrupts and then presenting the interrupt to the processor on a predetermined cycle count. Figure 13 shows a more detailed block diagram of the interrupt synchronization logic 65 of Figure 2. Each CPU
contains a distributor 135 which captures the external interrupt from the line 69 or 70 coming from the modules 14 or 15; this capture occurs on a predetermined cycle ` count, e.g., at count~ as signalled on an input line CC4 from the counter 71. The '1 captured interrupt is distributed to the other two CPUs via the inter-CPU bus 18.
These distributed interrupts are called pending interrupts. There are three pending interrupts, one from each CPU 11, 12 and 13. A voter circuit 136 captures the pending interrupts and perforrns a vote to verify that all of the CPUs did receive the ; external interrupt request. On a predetermined cycle count (detected from the cycle counter 71), in this exampb cycle-8 received by input line CC-8, the interrupt voter 136 presents the interrupt to the interrupt pin on its respective microprocessor 40 via line 137 and control bus 55 and 43. Since the cycle count that is used to present îhe interrupt is predetermined, all of the microptocessors 40 vill receive the interrupt on the same cycle count and thus the interrupt will have been synchronized to virtual time.

.:~
~ 38 ..

., . .
.~:

: .
. -.
. , .

- . .

Figure 14 shows the sequence of events for synchronizing interrupts to virtual time. The rows labeled CPU-A, CPU-B, and CPU-C indicatc the cycle count in counter 71 of each CPU at a point in real time. The rows labeled IRQ A PEND, IRQ B PEND, and IRQ C PEND indicate the state of the interrupt pending bits coupled via the inter-CPU bus 18 to the input of the voters 136 (a one signifies that the pending bit is set). The rows labeled IRQ A, IRQ B, and IRQ C indicate the state of the interrupt input pin on the microprocessor 40 (tbe signals on lines 137), where a one sigIufies that an interrupt is present at the input pin. In Figure 14, the external interrupt (EX IRQ) is asserted on line 69 at to~ If the interrupt distributor 135 captures and then distributes the interrupt to the inter-CPU bus 18 on cycle count 4, then IRQ C PEND will go active at tl, IRQ B PEND will go active at t2, and IRQ A PEND will go active at t4. If the interrupt voter 136 captures and then votes the interrupt pending bits on cycle count 8, then IRQ C will go active at t5, IRQ B
will go active at t6, and IRQ-A will go active at t8. Tbe result is that the interrupts were presented to the CPUs at different points in real time but at the same point in virtual time (i.e. cycle count 8).

Figure 15 illustrates a scenario which requires the algorithm presented in Figure 14 to be modified. Note that the cycle counter 71 is here represented by a modulo 8 counter. The external interrupt (EX IRQ) is asserted at time t3, and the ~ 20 interrupt distributor 13S captures and then distributes the interrupt to the inter-CPU
: bus 18 on cycle count 4. Since CPU-B and CPU-C have executed cycle count 4 t'''`' before time t3, their interrupt distributor does not capture the external interrupt.
CPU-A, however, executes cycle count 4 after tirne t3. The result is that CPU-A
l ~ captures and distributes the external interrupt at time t4. But if the interrupt voter l Y 2S 136 captures and votes the interrupt pending bits on cycle 7, the interrupt voter on CPU-A captures the IRQ A PEND signal at time t~, when the two other interrupt pending bits are not set. The interrupt voter 136 on CPU-A recognizes that not all of the CPUs have distributed the external interrupt and thus places the captured interrupt pending bit in a holding register 138. The interrupt voters 136 on CPU-B
and CPU-C capture the single interrupt pending bit at times t5 and t4 respectively.
Like the interrupt voter on CPU-A, the voters recognize that not all of the interrupt , ~ .

,.. .

!`
,;
..
.
s . . . . .. . ~ .
... .. . . .
~:.: ` .. .:
... . . . .
~;
- :

~003337 pending bits are set, and thus the singlc interrupt pending bit that is set is placed into the holding register 138. When the cycle counter 71 on each CPU reaches a cycle count of 7, the counter rolls over and begins counting at cycle count 0. Since the external interrupt is still asserted, the interrupt distributor 13S on CPU-B and CPU-C
S will capture the external interrupt at times tlo and t9 respectively. These times correspond to when the cycle co~mt becomes equal to 4. At time tl2, the interrupt voter on CPU-C captures the interrupt pending bits on the inter-CPU bus 18. The voter 136 determines that all of the CPUs did capture and distribute the external interrupt and thus presents the interrupt to the processor chip 40. At times tl33 and t~5, the interrupt voters 136 on CPU-B and CPU-A capture the interrupt pending bits and then presents the interrupt to the processor chip 40. The result is that all of the processor chips received the external interrupt request at identical instructions, and the information saved in the holding registers is not needed.
, ;
Holding Register:

lS In the interrupt scenario presented above with reference to Figure 15, the voter 136 uses a holding register 138 to save somc state information. In particular, the : saved state was that some, but not all, of the CPUs captured and distributed an external interrupt. If the system does not have any faults (as was the situation in Figure 15) then this state information is not nccessary because, as shown in the' 20 previous example, extemal interrupts can be synchronized to virtual dme without the il use of the holding register 138. The algorithm is that the interrupt voter 136 captures `~ and votes the interrupt pending biS on a predeterrnined cyclc count. When all of the , interrupt pending bits are asserted, then the interrupt is presented to the processor p chip 40 on the predetermined cycle count. In the example of Figure 15, the . 25 interrupts were voted on cycle count 7.
:
Referring to Figure 15, if CPU-C fails and the failure mode is such that the interrupt distributor 135 does not function correctly, then if the interrupt voters 136 . waited until all of the interrupt pending bits were set before presenting the interrupt ';`
"' , ., ., .. : ,'. :

;:
.. .~ . , .
~, . .

to the processor chip 40, the result would be that the interrupt would never getpresented. Thus, a single fault on a single CPU renders the entire interrupt chain on all of the CPUs inoperable.

The holding register 138 provides a mechanism for the voter 136 to know that the last interrupt vote cycle captured at least one, but not all, of the interrùpt pending bits. The interrupt vote cycle occurs on the cycle count that the interrupt voter captures and votes the interrupt pending bits. There arc only two scenarios that result in some of the interrupt pending bits being set. One is the scenario presented in reference to Figure 15 in which the extemal interrupt is asserted before the interrupt distribution cycle on some of the CPUs but after the interrupt distribution cycle on other CPUs. In the second scenario, at least one of the CPUs fails in amanner that disables the interrupt distributor. If the reason that only some of the interrupt pending bits are set at the interrupt vote cycle is case one scenario, then the interrupt voter is guaranteed that all of the interrupt pending bits will bc set on the next interrupt vote cycle. Therefore, if the interrupt voter discovers that the holding : register has been set and not all of the interrupt pending bits are set, then an error must exist on one or more of the CPUs. This assumes that the holding register 138 of each CPU gets cleared when an interrupt is serviced, so that the state of theholding register does not represent stale state on the interrupt pending bits. In the case of an error, the interrupt voter 136 can present the interrupt to the processor chip 40 and simultaneously indicate that an crror has been detected in the interrupt synchronization logic.

The interrupt voter 136 does not actualb do any voting but instead merely checks the state of the interrupt pending bits and the holding register 137 to 2S deterrnine whether or not to present an interrupt to the processor chip 40 and whether or not to indicate an error in the interrupt logic.
~"

Modulo Cycle Counters:

'`''"
,i:
~.
~.,, . . .
~ - .

;.. . - , . . : - .
. - . . :~ -. . - , . .. ..

., , ~ ' ' :' The interrupt synchronization example of Figure 15 represented the interrupt cycle counter 71 as a modulo N counter (e.g., a modulo 8 counter). Using a modulo N cycle counter simplified the description of the interrupt voting algorithm by allowing the concept of an interrupt vote cycle. With a modulo N cycle counter, the interrupt S vote cycle can be described as a single cycle count which lies between 0 and N-1 where N is the modulo of the cycle counter. Whatever value of cycle counter is chosen for the interrupt vote cycle, that cycle count is guaranteed to occur every N
cycle counts; as illustrated in Figure 15 for a modulo 8 counter, every eight counts an interrupt vote cycle occurs. The interrupt vote cycle is used here merely to illustrate the periodic nature of a modulo N cycle counter. Any event that is keyed to a particular cycle count of a modulo N cyde counter is guaranteed to occur every Ncycle counts. Obviously, an infinite (i.e., non-repeating counter 71) couldn't be used.

A value of N is chosen to maximize system parameters that have a positive effect on the system and to minimize system parameters that have a negative effect on the system. Some of such effects are~- developed empirically. First, some of the parameters will be described; Cv and Cd are the interrupt vote cycle and the interrupt distribution cycle respectively (in the circuit of Figure 13 these are the inputs CC-8 and CC-4, respectively). The value of Cv and Cd must lie in the range between ` O and N-1 where N is the modulo of the cycle counter. D"",~ is the maximum amount of cycle count drift between the three processors CPU-A, -B and -C that can be tolerated by the synchronization logic. The processor drift is determined by taking a snapshot of the cycle counter 71 from each CPU at a point in real time. The drift is calculated by subtracting the cycle count of the slowest CPU from the cycle count of the fastest CPU, performed as modulo N subtraction. The value of D",,,~ is ` i 2S describcd as a function of N and the values of Cv and Cd.
;~
.
i ~ First, D.,~" will be defined as a function of the difference Cv-Cd, where the subtraction operation is performed as modulo N subtraction. This allows us to choose ~, values of Cv and Cd that maLunize D"", Consider the scenario in Figure 16.
Suppose that Cd-8 and Cv-9. From Figure 16 the processor drift can be calculatedto be D,~ 4. The external interrupt on line 69 is asserted at time t4. In this case, .
.
.
.
~.
;.. I .. . . .

~.. ~ ,,. ' ~ -, :- .
~ ;. . . : : , ~' CPU-B will capture and distribute the interrupt at time t5. CPU-B will then capture and vote tbe interrupt pending bits at time t6. This scenario is inconsistent with the interrupt synchronization algorithm presented earlier because CPU-B executes itsinterrupt vote cycle before CPU-A has performed the interrupt distribution cycle. The flaw with this scenario is that the processors have drifted further apart than the difference between Cv and Cd. The relationship can be formally written as Equation (1) Cv - Cd ~ D~ - e where e is the time needed for the interrupt pending bits to propagate on the inter-CPU bus 18. In previous examples, e bas been assumed to be zero. Since wall-clock time has been quantized in clock cycle (Run cycle) increments, e can also be quantized. Thus the equation becomes Equation (2) Cv- Cd < D",,~ - 1 where D"~" is expressed as an integer number of cycle counts.

Next, the max~mum drift can be described as a function of N. Figure 17 lS illustrates a scenario in which N=4 and the processor drift D=3. Suppose that Cd=O. The subscripts on cycle count 0 of each processor denote the quotient part(Q) of the instruction cycle count. Since the cycle count is now represented in modulo N, the value of the cycle counter is the remainder portion of I/N where I is the number of instructions that have been executed since time tO. The Q of the instruction cycle count is the integer portion of l/N. If the external interrupt is asserted at time t3, then CPU-A will capture and distribute the interrupt at time t"
and CPU-B will execute its interrupt distribution cycle at time t5. This presents a problem because the interrupt distribution cycle for CPU-A has Qs1 and the interrupt distribution cycle for CPU-B has Q-2. The synchronization logic will continue as if there are no problems and will thus present the interrupt to the processors on equal cycle counts. But the interrupt will be presented to the processors on different instructions because the Q of each processor is different. The relationship of D",~" as a function of N is therefore ; Equation (3) N/2 > D, . .
.

` 43 .

. ~ ' Z0~3337 where 1~ is an even number and Dmax is expressed as an integer number of cycle counts. (These equations 2 and 3 can be shown to be both equivalent to the Nyquist theorem in sampling theory.) Combining equations 2 and 3 gives Equation (4) Cv - Cd < N/2 - 1 S which allows optimum values of Cv and Cd to be chosen for a given value o~ N.

All of the above equations suggest that N should be as large as possible. The only factor that tries to drive N to a small number is interrupt latency. Interrupt latency is the time interval between the assertion of the external interrupt on line 69 and the presentation of the interrupt to the microprocessor chip on line 137. Which processor should be used to determine the interrupt latency is not a clear-cut choice.
The three microprocessors will operate at different speeds because of the slightdifferences in the crystal oscillators in clock sources 17 and other factors. There will be a fastest processor, a slowest processor, and the other processor. Defining the interrupt latency with respect to the slowest processor is reasonable because the performance of system is ultimately determined by the performance of the slowestprocessor. The maximum interrupt latency is Equation (5) LmU = 2N -1 where LmU is the maximum interrupt latency expressed in cycle counts. The maximum interrupt latency occurs when the external interrupt is asserted after the interrupt distribution cycle Cd of the fastest processor but before the interrupt : distribution cycle Cd of the slowest processor. The calculation of the average interrupt latency L,Ve is more complicated because it depends on the probability that the çxternal interrupt occurs aftcr the interrupt distribution cycle of the fastest ; processor and before thc interrupt distribution cycle of the slowest processor. This probabiliq depends on the drift between the processors which in turn is determined by a number of external factors. If we assume that these probabilities are zero, then the average lateng may be expressed as Equation (6) L", = N/2 + (Cv - Cd) Using these relationships, values of N, Cv, and Cd are chosen using the system requirements for D,"," and interrupt lateng. For exampb, choosing N~128 and (Cv -Cd)=10, Lho=74 or about 4.4 microsec (with no stall gcles). Using the preferred ' :
: 44 .."`
~.~

~ . . ' :
.. .
~ .
~, . .
~ ,. :
;' embodiment where a four bit (four binary stage) counter 71a is used as the interrupt synch counter, and the distribute and vote outputs are at CC4 and CC-8 as discussed, it is scen that N=16, Cv=8 and Cd=4, so L,~,~=16/2 +(8-4) = 12-cycles or 0.7 rnicroscc.

Refresh Control for Local Memory:

The refresh counter 72 counts non-stall cycles (not machine cycles) just as the - counters 71 and 71a count. The object is that the refresh cycles will be introduced for each CPU at the sarne cycle count, measured in virtual time rather than real time.
Preferably, each one of the CPUs will interpose a refresh cycle at the same point in the instruction strearn as the other two. The DRAMs in local memory 16 must be refreshed on a 512 cycles per 8-msec. schedule just as mentioned abovc regarding the DRAMs 104 of the global memory. Thus, the counter 72 could issue a refresh command to the DRAMs 16 once every 15-microsec., addressing one row of 512, so the refresh specification would be satisfied; if a memory operation was requested during refresh then a 8usy response would result until refresh was finished. Butletting each CPU handle its own local memory refresh in real time independently of the others could cause the CPUs to get out of synch, and so additional control is needed. For example, if refresh mode is entered just as a divide operation is beginning, then timing is such that one CPU could take two clocks longer than others.
Or, if a non-interruptable sequence was entered by a faster CPU then the others went into refresh before entering this routine, the CPUs could walk away from one another. However, using the gcle counter 71 (instead of real time) to avoid some of these problems means that stall cycles are not countcd, and if a loop is enteredcausing man~ stalls (some can cause a 7-to-1 stall-to-run ratio) then the rcfresh specificadon is not met unless the period is decreased substantially from the 15-microsec figure, but that would degrade performance. For this reason, stall cycles are also counted in a second counter 72a, seen in Figure 2, and every time this counter reaches the same number as that counted in the refresh counter 72, an additionalrefresh cycle is introduced. For example, the refresh counter 72 counts 28 or 256 Run .,.

' ::.
. s .~

!,,: . . : . .
~' '" ~ ' '' .' ':. ' " ' ': '. ' . ' :

~003337 cycles, in step with the counter 71, and when it overflows a re&esh is signalled via control bus 43. Meanwhile, counter 72a counts 28 stall cycles (responsive ~o theRUN# signal and clock 17), and every time it over~ows a second counter 72b is incremented (counter 72b may be merely bits 9-to-11 for the eight-bit counter 72a), so S when a refresh mode is finally entered the CPU does a number of additional refreshes indicated by the number in the counter register 72b. Thus, if a long period of stall-intensive execution is encountered, the average number of re&eshes will stay in the one per 15-microsec range, even if up to 7x256 stall cycles are interposed, because when finally going into a refresh mode the number of rows re&eshed will catch up to the nominal re&esh rate, yet there is no degradation of performance by arbitrarily shortening the re&esh cycle.
-'' Memory Management:
..~
The CPUs 11, 12 and 13 of Figures 1-3 have memory space organized as illustrated in Figurc 18. Using the example that the local memory 16 is 8-MByte and the global memory 14 or 15 is 32-MByte, note that the local memory 16 is part of the same continuous zero-to-40M map of CPU memory access space, rather than being a cache or a separate memory space; realizing that the 0-8M section is triplicated (in the three CPU modules), and the 840M section is duplicated, nevcrtheless logically there is mcrely a single 0-40M physical address space. An address over 8-MByte on bus 54 causes the bus interface 56 to make a request to the memory modules 14 and 15, but an address under 8-MByte will access the local memory 16 within the CPU
module itself. Performance is improved by placing more of the memory used by theapplicadons being executed in local memory 16, and so as memory chips are available in higher dcnsides at lower cost and higher speeds, additional local memory will be added, as well as additional global memory. For example, the local memory might be 32~MByte and the global memory 128-MByte. On the other hand, if a very minimum-cost system is needed, and performance is not a major determining factor, thc system ....
~ can be operated with no local memory, all main memory being in the global memory ~ .:
1`'~``.
' ~ 46 ,, ~,' .

'' - `' '' '''' - .

~ .
~., 1-`'': : .

Z0~)3337 area (in memory modules 14 and 15), although the performance penalty is high ~orsuch a configuration.

The content of local memory portion 141 of the map of Figure 18 is identical in the three CPUs 11, 12 and 13. Likewise, the two memory modules 14 and 15 contain identically the same data in their space 142 at any given instant. Within the local memory portion 141 is stored the kernel 143 (code) for the Unix operating system, and this area is physically mapped within a fixed portion of the local memory 16 of each CPU. Likewise, kernel data is assigned a fixed area 144 in each localmemory 16; except upon boot-up, these blocks do not get swapped to or from global memory or disk. Another portion 145 of local memory 16 is employed for user prograrn (and data) pages, which are swapped to area 146 of the global memory 14and 15 under control of the operating system. The global memory area 142 is usedas a staging area for user pages in area 146, and also as a disk buffer in an area 147;
if the CPUs are executing code which performs a write of a block of data or codefrom local memory 16 to disk 148, then the sequence is to always write to a diskbuffer area 147 instead because the time to copy to area 147 is negligible compared to the ti~ne to copy directly to the I/O processor 26 and 27 and thus via I/O
controller 30 to disk 148. Then, while the CPUs proceed to execute other code, the write-to-disk operation is done, transparent to the CPUs, to move thc block from area 147 to disk 148. In a like manner, the global memory area 146 is mapped to include an I/O staging 149 area, for similar treatment of I/O accesses other than disk (e.g., video).
:
The physical memory map of Figure 18 is correlated with the virtual memory ~ management system of the processor 40 in each CPU. Figure 19 illustrates the vinual ; ~ 25 address map of the R2000 processor chip uscd in the example embodiment, although it is understood that other microprocessor chips supporting virtual memory manage-ment with paging and a protection mechanism would provide corresponding features.

In Figure 19, two separate 2-GByte virtual address spaces 150 and 151 are illustrated; the processor 40 operates in one of two modes, user modc and kernel ., .

. .

, , .
. . ~
..

.'.. : .

.:

: . . ~, ~, . ~ . .. - .
.

mode. The processor can only access the area 150 in the user mode, or can accessboth the areas 150 and 151 in the kernel mode. The kernel mode is analogous to the supenrisory mode provided in many machines. The processor 40 is configured to operate normally in the user mode until an exception is detected forcing it into the kernel mode, where it remains until a restore from exception (RFE) instruction is executed. The manner in which the memory addresses are translated or mapped depends upon the operating mode of the microprocessor, which is defined by a bit in a status register. When in the user mode, a single, uniform virtual address space 150 referred to as "kuseg" of 2-GByte size is available. Each virtual address is also extended with a 6-bit process identifier (PID) field to form unique virtual addresses for up to sixlty-four user processes. All references to this segment 150 in user mode are mapped through the TLB 83, an~ use of the caches 144 and 145 is determined by bit settings for each page entry in the TLB entries; i.e., some pages may be cachable and some not as specified by the programmer.

lS When in the kernel mode, the virtual address space includes both the areas - 150 and 151 of Figure 19j and this space has four separate segments kuseg 1SO, ksegO
152, ksegl 153 and kseg2 154. The kuseg 150 segment for the kernel mode is 2-GByte in size, coincident with the "kuseg" of the user mode, so when in the kernel mode the processor treats references to shis segment just like user mode references, thus streamlining kernel access to user data. The kuseg 150 is used to hold user code and data, but the operating system often needs to reference this same code or data.
The ksegO area 152 is a S12-MByte kemel physical address space dircct-mapped onto the first 512-MBytes of physical address space, and is cached but does not use the TLB 83; this segment is used for kemel executable code and some kernel data, and is 2S represented by the area 143 of Figure 18 in local memory 16. The ksegl area lS3 is ' also directly mapped into the first 512-MByte of physical address space, the same as `` ksegO, and is uncached and uses no TLB entries. Ksegl differs from ksegO only in that it is uncached. Ksegl is used by the operating system for 1/0 registers, ROM
code and disk buffers, and so corresponds to areas 147 and 149 of the physical map of Figure 18. The kseg2 area 154 is a 1-GByte space which, like kuse& uses TLB 83 entries to map virtual addresses to arbitrary physical ones, with or without caching.

.

' ..

~ - .
,~, . . . . .
~, .
. . - . . . . .
- .

. .' - ~ .

~003337 This kseg2 area differs from the kuseg area 150 only in that it is not accessible in the user mode, but instead only in the kerncl mode. The operating system uses kseg2 for stacks and per-process data that must remap on context switches, for user page tables (memory map~, and for some dynarnically-allocated data areas. Kseg2 allows selective caching and mapping on a per page basis, rather than requiring an all-or-nothingapproach.

The 32-bit virtual addresses generated in the registers 76 or PC 80 of the microprocessor chip and output on the bus 84 are represented in Figure 20, where it is seen that bits 0-11 are the offset used unconditionally as the low-order 12-bits of the address on bus 42 of Figure 3, while bits 12-31 are the VPN or virtual page number in which bits 29-31 select between kuseg, ksegO, ksegl and kseg2. The process identifier PID for the currently-executing process is stored in a register also accessible by the TLB. The 64-bit TLB entries are represented in Figure 20 as well, where it is seen that the 20-bit VPN from the virtual address is compared to the 20-bit VPN field located in bits 44-63 of the 64-bit entry, while at the same time the PID is compared to bits 38-43; if a match is found in any of the sixty-four 64-bit TLB
entries, the page frame number PFN at bits 12-31 of the matched entry is used as the output via busses 82 and 42 of Figure 3 (assuming othcr criteria are met). Otherone-bit values in a TLB entry include N, D, V and G. N is the non-cachable indicator, and if set the page is non-cachable and the processor directly accesses local memory or global memory instead of first accessing the cache 44 or 4S. D is a write-protect bit, and if set means that the location is "dirty" and therefore writable, but if zerG a write operation causes a trap. The V bit mcans valid if set, and allows the l~LB cntries to be cleared by merely resetting the valid bits; this V bit is used in the 2S page-swapping arrangement of this system to indicate whether a page is in local or ~; global memory. The G bit is to allow global accesses which ignore the PID match requirement for a valid TLB translation; in kseg2 this allows the kernel to access all ' mapped data without regard for PID.
~: 1 The device controllers 30 cannot do DMA into local memory 16 directly, and so the global memory is used as a staging area for DMA type block transfers, .. ..

. .
".
~.' ~' '. .
.
~, : . , . . -~ ,;

. ~ , . .

;:

-, . . .

typically from disk 148 or the like. The CPUs can pcrform operations directly at the controllers 30, to initiatc or actually control operations by the controllers (i.e., programmed I/O), but thc controllers 30 cannot do DMA except to global memory;
the controllers 30 can become the VMEbus (bus 28) master and through the I/O
S processor 26 or 27 do reads or writes directly to global memory in the memory modules 14 and 15.

Page swapping between global and local memories (and disk) is initiated either by a page fault or by an aging process. A page fault occurs when a process is executing and attempts to execute from or access a page that is in global memory or on disk; the TLB 83 will show a miss and a trap will result, so low level trap code in the kernel will show the location of the page, and a routine will be entered to initiate a page swap. If the page needed is in global memory, a series of commands are sent to the DMA controller 74 to write the least-recently-used pagc from local memory to global memory and to read the needed page from global to local. If the page is on disk, commands and addresses (sectors) are written to the controller 30 from the CPU
to go to disk and acquire the page, then the process which made the memory reference is suspended. When the disk controller has found the data and is ready to send it, an interrupt is signalled which will be used by the memory modules (not, reaching the CPUs) to allow the disk controller to bcgin a DMA to global memory to write the page into global memory, and when finished the CPU is interrupted to begin a block transfer undcr control of DMA controller 74 to swap a least used page from local to global and read the needed page to local. Then, the original process is made runnable again, state is restored, and the original memory reference will again occur, finding the needed page in local memory. The other mechanism to initiate page swapping is an aging routine by which the operating system periodically goes through the pages in local memory marking them as to whether or not each page has been used recently, and those that have not are subject to be pushed out to global i; memory. A task switch does not itself initiate page swappin& but instead as the new ! ~ task begins to produce page faults pages will be swapped as needed, and the ! 30 candidates for swapping out are those not recently used.

~', 50 L

.~
!`
, ~.
, . . . .
... ,- ~. . . -.
,.
` .
.
. `
':

200333~7 If a memory reference is made and a TLB miss is shown, but thc pagc table lookup rcsulting from thc TLB miss exception shows the page is in local mcmory, thcn a TLB entry is made to show this pagc to bc in local memory. That is, the process takes an exception when the TLB miss occurs, gocs to thc pagc tablcs (in thc kcrnel data section), finds thc table cntry, writes to TLB, then thc proccss is allowcd to proceed. But if the memory reference shows a TLB miss, and the page tables show the corresponding physical address is in global memory (over 8M physical address), the TLB entry is made for this page, and when the proccss rcsumes it will find thc pagc cntry in thc TLB as before; yet another exception is taken because the valid bit will bc zcro, indicating thc page is physically not in local mcmory, so this timc thc exception will enter a routinc to swap thc pagc from global to local and validatc the TLB entry, so exccution can thcn proceed. In the third situation, if the page tables show address for the memory reference is on disk, not in local or global memory, then the system operates as indicated above, i.e., the process is put off the run queue and put in the sleep queue, a disk request is made, and when the disk has transferred the page to global memory and signalled a command-complete interrupt, then the page is swapped from global to local, and the TLB updated, then the process can execute again.

Private Memory:

Although the memory modules 14 and 15 store the same data at the same locations, and all three CPUs 11, 12 and 13 have equal access to these memory modules, there is a small area of the memory assigned under software control as a private memory in each one of the memory modules. For example, as illustrated inFigure 21, an area 155 of the map of the memory module locations is designated the 2S private memory area, and is writable only when the CPUs issue a "private memory write" command on bus S9. In an example embodiment, the private memory area 155 is a 4K page starting at the address contained in a register 156 in the bus interface 56 of each one of the CPU modules; this starting address can be changed under software control by writing to this register 156 by the CPU. The private memory area :``
: `~

,: - . . . ~
~ , . . . .

155 is further divided between the three CPUs; only CPU-A can write tO area 155a, CPU-B to area lSSb, and CPU-C to area 155c. One of thc command signals in bus 57 is set by the bus interface 56 to inform the memory modules 14 and 15 that the operation is a private write, and this is set in response to the address generated by S the processor 40 from a Store instruction; bits of the address (and a Write command) are detected by a decoder 157 in the bus interface (which compares bus addresses to the contents of register 156) and used to generate the "privatc memory write"
command for bus 57. In the memory module, when a write command is detected in the registers 94, 95 and 96, and the addresses and commands are all voted good (i.e., in agreement) by the vote circuit 100, then the control circuit 100 allows the data from only one of the CPUs to pass through to the bus 101, this one being determined by two bits of the address from the CPUs. During this private write, all three CPUs present thc same address on their bus 57 but different data on their bus 58 (thedifferent data is some state unique to the CPU, for example). The memory modulesvote the addresses and commands, and select data from only one CPU based upon part of the address field seen on the address bus. To allow the CPUs to vote some .. data, all three CPUs will do three private writes (there will be three writes on the busses 21, 22 and 23) of some state information unique to a CPU, into both memory modules 14 and 15. During each write, each CPU sends its unique data, but only one is accepted each time. So, the software sequence executed by all three CPUs is (1) Store (to location lSSa), (2) Store (to location 155b), (3) Store (to location 155c).
8ut data from only one CPU is actually written each time, and the data is not voted (because it is or could be diffcrent and could show a fault if voted). Then, the CPUs can vote the data by having all three CPUs read all three of the locations 155a 155b and 155c, and by software compare this data This type of operation is used in diagnostics, for example, or in interrupts to vote the cause register data The pnvate-write mechanism is used in fault detection and recovery. For example, if the CPUs detect a bus error upon malcing a memory read request, such as a memory module 14 or lS returning bad status on lines 33-1 or 33-2. At this point . ., j 30 a CPU doesn't know if the other CPUs received the same status from the memory i module; the CPU could be faulty or its status detection circuit faulty, or, as indicated, ~ `
:
~. . .
,.
';
~...

~ .
~, . . . : .
. ..

,": , , . -, -,; " ' .
.

the memory could be faulty. So, to isolate the fault, when the bus fault routinemendoned above is entered, all three CPUs do a private write of the status information they just received from the memory modules in the preceding read attempt. Then all three CPUs read what the others have written, and compare it S with their own memory status information. If they all agree, then the memory module is voted off-line. If not, and one CPU shows bad status for a memory module but the others show good status, then that CPU is voted off-line.

Fault-Tolerant Power Supply:

Referring now to Figure 22, the system of the preferred embodiment may use a hult-tolerant power supply which provides the capability for on-line replacement of failed power suppb modules, as well as on-line replacement of CPU modules, memory modules, I/O processor modules, I/O controllers and disk modules as discussed above.
In the circuit of Figure æ, an a/c power line 160 is connected directly to a power distribution unit 161 that provides power line filtering, transient suppressors, and a circuit breaker to protect against short circuits. To protect against a/c power line failure, redundant battery packs 162 and 163 provide 4-1/2 minutes of full system power so that orderly system shutdown can be accomplished. Only one of the two battery packs 162 or 163 is required to be operative to safely shut the system down.
:' The power subsystem has two identical AC to DC bulk power supplies 164 and 165 which exhibit high power factor and energize a pair of 36-volt DC distribution busses 166 and 167. The system can remain operational with one of the bulk powersupplies 164 or 16S operadonal.
, ~
Four separate power distribution busses are included in these busses 166 and 167. The bulk supply 164 drives a power bus 166-1, 167-1, while the bulk supply 165 2S drives power bus 166-2, 167-2. The battery pack 162 drives bus 166-3, 167-3, and is itself recharged from both 166-1 and 166-2. The battery pack 163 drives bus 166-3, , S3 , ~, .
:

~ , . . .
. . , ~ . , ~ .
:, . .
'-' : ~ , .'~ . ' . . .
., . :
. :,, . : .
. . .
-. . :. . :.

167-3 and is recharged from busses 166-1 and 167-2. The three CPUs 11, 12 and 13are driven from different combinations of these four distribution busses.

A number of DC-to-DC converters 168 connected to these 36-v busses 166 and 167 are used to individually power the CPU modules 11, 12 and 13, the memory modules 14 and 15, the I/O processors 26 and 27, and the I/O controllers 30. Thebulk power supplies 164 and 165 also power the three system fans 169, and battery chargers for the battery packs 162 and 163. By having these separate DC-to-DC
converters for each system component, failure of one converter does not result in system shutdown, but instead the system will continue under one of its &ilure recovery modes discussed above, and the failed power supply component can be replaced while the system is operating.

The power system can be shut down by either a manual switch (with standby and off functions) or under software control from a maintenance and diagnostic processor 170 which automatically defaults to the power-on state in the event of a maintenance and diagnostic power failure.

While the invention has been described with reference to a specific embodi-ment, the description is not meant to be construed in a limiting sense. Various modifications of the disclosed embodiment, as well as other embodiments of the invention, will be apparent to persorls skilled in the art upon reference to this description. It is therefore contemplated that the appended claims will cover any such modifications or embodiments as fall within the true scope of the invention.

.
' ~;. 54 ~`
,, ,:

, .. . . . . .

-~ , ` ' , ' '

Claims (38)

1. A fault-tolerant computer system, comprising:
a) first, second and third CPUs of substantially identical configuration, said first, second and third CPUs executing substantially the same instruction stream;
b) first and second memory modules of substantially identical configuration, said first and second memory modules storing substantially the same data;
c) busses coupling each of the first, second and third CPUs individually to each of said first and second memory modules whereby said first, second and third CPUs access said first and second memory modules separately and in duplicate;
d) a first input/output bus coupled to said first memory module and a second input/output bus coupled to said second memory module; and e) a first input/output processor coupled to both said first and second input/output busses, and a second input/output processor coupled to both said first and second input/output busses.

2. A system according to claim 1 wherein said first, second and third CPUs are loosely synchronized upon the event of a memory reference.

3. A system according to claim 2 wherein said first, second and third CPUs are loosely synchronized upon the event of a memory reference by detecting an access by said CPUs to said first and second memory modules and stalling any CPUs for which the access occurs earlier to wait until the last one of said CPUs executes said access, then allowing the access to occur.

4. A system according to claim 1 wherein said first, second and third CPUs are operating on independent clocks so that said execution is asynchronous.

5. A system according to claim 1 wherein either of said memory modules is designated as a primary memory and the other is designated as back-up, and wherein write operations by the CPUs are executed in both of said memory modules but in read operations said CPUs receive data from only the primary memory module; and wherein the back-up memory module may be designated as the primary memory and the primary may be designated as the back-up, at any time.

6. A system according to claim 1 including at least one additional input/output processor connected to both said first and second input/output busses.

7. A system according to claim 1 wherein said memory modules include means for voting said memory references, and wherein data is voted only for memory writes in said means for voting said memory references, and addresses and commands are voted for both memory reads and memory writes.

8. A system according to claim 7 wherein each of said memory modules includes a random access memory and said random access memory is accessed in parallel with operation of said means for voting.

9. A system according to claim 1 wherein when one of said first, second and third CPUs fails, said one is placed off-line and the remaining two continue to execute said instruction stream.

10. A method of operating a redundant computer system, comprising the steps of:
a) executing the same instruction stream in each of first, second and third individually-clocked CPUs;
b) storing identical data in first and second memory modules;
c) executing memory references by each of the first, second and third CPUs individually to each of said first and second memory modules; and d) performing store or recall operations in said first and second memory modules in response to said memory references when said memory references have been received from said first, second and third CPUs.

11. A method according to claim 10 including the steps of:

e) executing input/output functions separately to a first input/output bus coupled to said first memory module and to a second input/output bus coupled to said second memory module; and f) processing said input/output functions in a first input/output processor connected to both said first and second input/output busses, and in a second input/output processor connected to both said first and second input/output busses.

12. A method according to claim 11 including the step of loosely synchroniz-ing said first, second and third CPUs by detecting synchronization upon the event of a memory reference and stalling any one of the CPUs leading another CPU so the event is executed by all three in synchronization.

13. A method according to claim 10 including the step of processing said input/output functions in at least one additional input/output processor connected to both said first and second input/output busses.

14. A method according to claim 11 including the step of voting said memory references and thereby synchronizing said first, second and third CPUs with one another.

15. A method according to claim 10 including the step of voting said memory accesses to both said first and second memory modules from said first, second and third CPUs; and including the step of accessing the memory of said each said first and second memory modules while at the same time performing said step of voting so that when the step of voting is completed the memory access can be immediately validated.

16. A computer system comprising:
a) first, second and third CPUs executing the same instruction stream, each of said CPUs having a separate memory access port;
b) first and second memory means having identical address spaces within the address range of said CPUs for storing duplicative data, each one of said first and second memory means having first, second and third input/output ports coupled tosaid memory access ports of said first, second and third CPUs, respectively;
c) each one of said first, second and third ports including voting means to compare information between the ports for accesses and to allow accesses to be completed only where at least two of the ports have the same such information.

17. A computer system according to claim 16 including first and second input/output busses, and wherein said first and second memory means each has a separate peripheral I/O port, and said first and second input/output busses are separately coupled to said I/O ports of said first and second memory means.

18. A computer system according to claim 16 including means for synchroniz-ing execution of the instruction stream in said CPUs by stalling execution of a memory reference until all three of said first, second and third CPUs are executing the exact same memory reference.

19. A system according to claim 16 wherein said information includes address and data information for write operations.

20. A system according to claim 16 wherein each one of said CPUs has a separate clock whereby said CPUs operate asynchronously, and wherein each one ofsaid first and second memory modules has a separate clock, whereby said CPUs andmemory modules all operate asynchronously.

21. A system according to claim 20 wherein said CPUs are loosely synchroniz-ed with one another by stalling leading CPUs until the slower CPUs have caught up, upon the event of a memory reference.

22. A method of operating a computer system comprising the steps of:
a) executing the same instruction stream in first, second and third CPUs;
b) generating memory accesses in each of said first, second and third CPUs at separate first, second and third memory access busses;

c) storing duplicative data in first and second memory means having substantially identical address spaces within the address range of said CPUs, including executing memory accesses to each one of said first and second memory means via said first, second and third memory busses;
d) voting each one of said memory accesses in said first and second memory means when received from said first, second and third memory access busses, said voting including comparing information representing said memory accesses; and (e) allowing said accesses to be completed only where at least two of said access busses present the same such information.

23. A method according to claim 22 including the step of synchronizing said first, second and third CPUs whereby said CPUs are substantially simultaneously executing the same instruction stream, and wherein said step of synchronizing said CPUs includes stalling execution of memory accesses until all three of the first, second and third CPUs is executing the same memory access at the same time.

24. A method according to claim 23 wherein said step of synchronizing also includes timing the implementation of external interrupts of the CPUs so that all three of the first, second and third CPUs are executing the same instruction at the time the interrupt is presented.

25. A method of operating a computer system comprising the steps of:
a) executing the same instruction stream in at least first and second processors;
b) generating remote accesses in each of said first and second processors, the remote accesses being directed to separate first and second access ports;
e) detecting each one of said remote accesses at said first and second access ports and stalling executing of instructions in said first and second processors until a remote access is detected at both the first and second access ports.

26. A method according to claim 25 wherein said step of executing is in first, second and third processors, and wherein said step of generating remote accesses is in said first, second and third processors.

27. A method according to claim 26 wherein said first and second access ports are in first and second modules operated asynchronous to said first, second and third processors; and wherein said first and second processors are asynchronous.

28. A method according to claim 25 including the step of storing duplicative data in first and second address spaces within the address range of said processors, and executing accesses to each one of said first and second address spaces via first and second access busses.

29. A method according to claim 25 including the step of synchronizing any interrupts executed by said first and second processors whereby said processors execute the same interrupt at the same instruction in said instruction stream.

30. A method according to claim 29 wherein said step of synchronizing said processors includes stalling execution of memory accesses until all three of the first, second and third processors are executing the same memory access at the same time.

31. A method according to claim 29 wherein said step of synchronizing also includes stalling execution of interrupts of the CPUs until all three of the first, second and third processors are executing the same interrupt at the same time.

32. A computer system comprising:
a) multiple CPUs;
b) redundant modules accessed by said CPUs via separate ports from each CPU, the modules voting requests received from the CPUs;
c) redundant I/O processors accessed by said CPUs via said ports and said modules;

d) a plurality of I/O devices each of which is accessed by one of said I/O processors as designated by data stored by said CPUs, said data being alterable by said CPUs to redesignate the I/O processor accessing each said I/O device.

33. A system according to claim 32 wherein said multiple CPUs execute the same instruction stream, and wherein said modules are memory modules.

34. A system according to claim 32 including means for detecting faults in said I/O processors, wherein said data stored by said CPUs is the addresses of registers in said I/O devices, and wherein said each one of said I/O processors is accessed by said CPUs via two of said modules.

35. A method of operating a computer system having multiple CPUs comprising the steps of:
a) accessing redundant modules by said multiple CPUs via separate ports from each CPU, including voting requests received from the CPUs by the modules;
b) accessing redundant I/O processors by said CPUs via said ports and said modules;
c) accessing a plurality of I/O devices via said I/O processors, each of the I/Odevices being accessed by one of said I/O processors as designated by data stored by said CPUs, said data being alterable by said CPUs to redesignate the I/O processor accessing each said I/O device.

36. A method according to claim 35 wherein said multiple CPUs execute the same instruction stream, and wherein said modules are memory modules.

37. A method according to claim 35 including the step of detecting faults in said I/O processors, and wherein said data stored by said CPUs is the addresses of registers in said I/O devices.

38. A method according to claim 35 wherein said each one of said I/O
processors is accessed by said CPUs via two of said modules.