CN103491018A - Load balancing method in large-scale behavior modeling application - Google Patents

Abstract

The invention discloses a load balancing method in large-scale behavior modeling application. The load balancing method comprises the following steps that two hash tables are established, wherein one hash table is used for storing all mock objects in a block, and the other hash table is used for storing all block messages in a domain; the hash tables are initialized, increment computation is conducted repeatedly on loads of all domains until the large-scale behavior modeling application is completed, a domain D after load increment calculating is generated, the domain D is divided to generate a domain set Dset, wherein the load of each domain D2 in the divided domain set Dset is T1, all the blocks are mutually connected, the domain D and adjacent domains in the divided domain set Dset are combined, the load of the combined domain D is T1, all the blocks in the domain D are mutually connected, the domain dispatch among all data processors is carried out, and the load balancing of the data processors is achieved. The load balancing method in the large-scale behavior modeling application is capable of solving the problem that load unbalancing can not be effectively removed in an existing method in the large-scale behavior modeling application.

Description

Load-balancing method in a kind of extensive Behavior modeling application

Technical field

The invention belongs to the teaching method technical field, more specifically, relate to the load-balancing method in a kind of extensive Behavior modeling application.

Background technology

Extensive Behavior modeling application is for the extensive autonomous individual group behavior in the simulating reality world, better learns and understands the phenomenon in real world.Such application is dispersed throughout physics, the fields such as biology and anthropology, for example: shoal of fish simulation, celestial body simulation and traffic simulation etc.Such as aspect traffic, can before a traffic construction, utilize the traffic simulation program, the jam situation of look-ahead road, reduce road and realize later traffic congestion situation.These application are with regular time interval operation.When each Fixed Time Interval, simulated object is collected the information of peripheral simulated object, processes and upgrades the operation of its state information for next time interval.For these extensive Behavior modeling programs of operation parallel on group system, the division of data is to have carried out in the mode in territory.That is to say, whole virtual space is divided into a plurality of territories, and all simulated object in territory are assigned to together and process.Like this just can improve data locality, the completing of faster procedure.

Yet, because simulated object in extensive Behavior modeling can group be carried out the position migration, causing simulated object number in above-mentioned said some territory is the several times decades of times even of simulated object number in its co-domain, causes serious load imbalance.And the simulated object number in each territory, along with the time is also changing, makes present method be difficult to effectively solve this load imbalance problem.

Summary of the invention

Above defect or Improvement requirement for prior art, the invention provides the load-balancing method in a kind of extensive Behavior modeling application, its purpose is to solve existing method in extensive Behavior modeling application and is difficult to the effectively technical problem of the load imbalance of elimination.

For achieving the above object, according to one aspect of the present invention, load-balancing method in a kind of extensive Behavior modeling application is provided, to be applied in a kind of load equalizer, the data of this load equalizer for moving its place data processor data management system, to realize the load balancing between data processor, the method comprises the following steps:

(1) set up two Hash tables, one of them is for depositing all simulated object of piece, and another is for depositing all block messages in territory, and this Hash table is carried out to initialization;

(2) repeat increment calculates the load in each territory, until extensive Behavior modeling application end, with the territory D after the load of generation incremental computations; This step comprises following sub-step:

(2-1) judge whether that the processed complete simulated object of all new rounds is all processed, if yes then enter step (2-8), otherwise enter step (2-2);

(2-2) obtain the new location information (O of the simulated object O after processed in each territory _x, O _y, O _z);

(2-3) according to the new location information of simulated object and calculate the numbering K of this simulated object place piece according to following formula _new:

$\left\{\begin{array}{c}x=\mathrm{Ox}/S_x\\ y=\mathrm{Oy}/S_y\\ z=\mathrm{Oz}/S_z\\ \mathrm{Knew}=x+y\×\mathrm{Nx}+z\×\mathrm{Nx}\×\mathrm{Ny}\end{array}\right.$

N wherein _x, N _yand N _zthe extensive Behavior modeling application simulation space x in user-defined edge, the number of blocks of y and z axle, S _x, S _yand S _zbe piece respectively along x, the size of y and z axle;

(2-4) the block number K of gained is calculated in judgement _newwhether identical with the block number at simulated object O place after processed, if different, enter step (2-5), otherwise return to step (2-1);

(2-5) by this simulated object from the old K that is numbered _oldpiece in take out, and this simulated object is inserted and is numbered K _newpiece in, and upgrade in the piece table and be numbered K _oldbe numbered K _newthe simulated object of piece count information and simulated object list information;

(2-6) judgement is numbered K _oldbe numbered K _newthe simulated object number change value Δ of piece | whether p| surpasses threshold value T, if surpass, just enters step (2-7), otherwise returns to step (2-1);

(2-7) incremental update is numbered K _oldbe numbered K _newthe load in territory, piece place and the load of this territory adjacent domains, then return to step (2-1);

(2-8) judge whether that all territories all are disposed, if enter step (5), otherwise enter step (2-9);

(2-9) whether the load that judges the territory D after all incremental computations loads is greater than T _max, if so, enter step (3), otherwise enter step (2-10);

(2-10) whether the load that judges the territory D after all incremental computations loads is less than T _min, if so, enter step (4), otherwise return to step (2-8);

(3) territory D is divided to generate territory set D _set, the territory set D after wherein dividing _setin the load of each territory D2 be T1, and piece wherein is all interconnective; This step comprises following sub-step:

(3-1) edge of newly-built territory D2，Cong territory D takes out a load and is less than the piece P1 of T1, and is inserted into territory D2;

(3-2) obtain all set of blocks P that are connected with piece P1 in the D of territory _set;

(3-3) from set of blocks P _setmiddle taking-up piece P2;

(3-4) after decision block P2 inserts territory D2, whether the load of D2 is less than T1, if so, enters step (3-5), otherwise enters step (3-7);

(3-5) piece P2 is inserted to D2, and upgrade the load of D2;

(3-6) obtain set of blocks adjacent with P2 in D, and it is incorporated to set P _set, then return to step (3-3);

(3-7) D2 is inserted to resultant field set D _set;

(3-8) whether the load that judges D still is greater than T _max, if so, return to step (3-1), otherwise return to step (2-8);

(4) by territory D and its territory set D after division _setin adjacent domains merged, the load of the territory D after wherein merging is T1, and the piece in the D of territory is all interconnective;

(5) carry out the domain scheduling between each data processor, realize the load balancing of data processor, wherein each territory of same data processor interconnects.

(6) notice tasks carrying device starts the execution of an extensive Behavior modeling application new round;

(7) judge that whether extensive Behavior modeling program new round calculating finishes, and if so, enters step (8), otherwise returns to step (7);

(8) judge whether extensive Behavior modeling program has reached user-defined termination condition, if so, finishes the operation of load equalizer, otherwise return to step (2).

Preferably, step (1) comprises following sub-step:

(1-1) set up a piece table and a territory table, wherein the piece table is for depositing all simulated object of piece, the corresponding data group of each list item, and comprise that four are listd: first lists the key assignments for memory block, lising for second is a chained list, the key assignments that is used for the adjacent block of memory block, the 3rd simulated object number of lising for memory block, lising for the 4th is a chained list, all simulated object information for memory block, the territory table is for depositing all block messages in territory, each list item is a corresponding data group also, and comprise that four are listd: first is listd for storing the key assignments in territory, list for storing the load in territory for second, list for storing the adjacent domains information in territory for the 3rd, lising for the 4th is a chained list, for storing the key assignments of all that territory comprises,

(1-2) the whole virtual space of extensive Behavior modeling application is divided into to the piece of fixed size, the information of all is filled out in the write-in block table, and be averagely allocated to a plurality of territories by all;

(1-3) load in all territories of initialization:

(1-4) notice tasks carrying device starts the execution of extensive Behavior modeling application.

Preferably, step (1-3) is to adopt following formula:

$\left\{\begin{array}{c}\mathrm{Cost}\left(D\right)=\underset{P\∈D}{\mathrm{\Σ}}\mathrm{Cost}\left(P\right)\\ \mathrm{Cost}\left(P\right)=\left|P\right|\×\underset{Q\∈\mathrm{Neighbors}\left(P\right)}{\mathrm{\Σ}}\left|Q\right|,\end{array}\right.$

Cost(D wherein) and the Cost(P) load of representative domain D and the load of the piece P in IncFlds D respectively, and | P| and | Q| means respectively the simulated object quantity that piece P and piece Q comprise, and Neighbors(P) means the adjacent block set of piece P.

Preferably, step (2-7) is to adopt following formula:

$\left\{\begin{array}{c}\mathrm{Cost}\left(D\right)=\mathrm{Cost}\left(D\right)+\underset{P\∈D}{\mathrm{\Σ}}\mathrm{\ΔCost}\left(P\right)\\ \mathrm{\ΔCost}\left(P\right)=\left|P\right|\×\mathrm{\ΔN}+\mathrm{\Δ}\left|P\right|\×N+\mathrm{\Δ}\left|P\right|\×\mathrm{\ΔN}\\ \mathrm{\ΔN}=\underset{Q\∈\mathrm{Neighbors}\left(P\right)}{\mathrm{\Σ}}\mathrm{\Δ}\left|Q\right|,\end{array}\right.$

Wherein Δ means the difference to dependent variable.

Preferably, step (4) comprises following sub-step:

(4-1) obtain first adjacent domains D1 of territory D;

(4-2) judge that whether D1Yu territory, territory D is at same data processor, and their load sum is less than threshold value T1, if so, enters step (4-3), otherwise enter step (4-4);

(4-3) territory D1 is merged in the D of territory, and the more load of neofield D, step (4-4) then entered;

(4-4) obtain the next adjacent domains D1 of D;

(4-5) judge that whether D1 is empty, if so, returns to step (2-8), otherwise enters step (4-2).

Preferably, step (5) comprises following sub-step:

(5-1) whether the load that judges current data processor, over threshold value Tt, if so, enters step (5-2), otherwise enters step (6);

(5-2) obtain a territory D of current data processor;

(5-3) obtain the data processor set W at all territory place adjacent with territory D;

(5-4) take out the data processor w1 of least-loaded in W;

(5-5) whether the load that judgement moves to territory D after data processor w1 is less than Tt, if so, enters step (5-6), otherwise enters step (5-7);

(5-6) territory D is moved to data processor w1 above, then enter step (5-1);

(5-7) obtain the data processor WL of least-loaded on all data processors, territory D is moved to this data processor WL above, and enter step (5-1).

In general, the above technical scheme of conceiving by the present invention compared with prior art, can obtain following beneficial effect:

1, load balancing, the cluster utilance is high, program operation speed is fast: because step (2) can solve task load, calculate inaccurate problem, step (3) and step (4) can be divided/merge extremely excessive/too small task of load and be solved extremely unbalanced problem of load, step (5) can be carried out load imbalance problem between data processor, therefore this method can be eliminated the load imbalance in extensive Behavior modeling application, improve the resource utilization of large-scale cluster system, accelerate the speed of simulation program.

2, run-time overhead is few: the load of obtaining each task that can increment due to step (2) solves all task loads and will recalculate the overhead issues of bringing at every turn, step (3-5) only needs on the node that extremely excessive/too small task of load was divided/merged to increment and the part task on the too much node of task quantity of only need to moving is very few to task quantity to solve virtual space all needs to repartition the overhead issues with the mass data migration brought far from it of twice of front and back division result at every turn, so the expense of this method is few.

The accompanying drawing explanation

Fig. 1 is the applied environment figure of the load-balancing method in the extensive Behavior modeling application of the present invention.

Fig. 2 is the flow chart of the load-balancing method in the extensive Behavior modeling application of the present invention.

Fig. 3 is the refinement flow chart of step in the inventive method (3).

Fig. 4 is the refinement flow chart of step in the inventive method (4).

Embodiment

In order to make purpose of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein, only in order to explain the present invention, is not intended to limit the present invention.In addition, below in each execution mode of described the present invention involved technical characterictic as long as form each other conflict, just can mutually not combine.

General thought of the present invention is, a kind of load-balancing algorithm of increment has been proposed, it can be by the load of each territory of effective real-time assessment (each territory is processed by a task of correspondence), carry out the division of automatic territory and merging on the basis of former division result, effectively guarantee that the load in each territory is much the same.When new one takes turns analog computation and starts, when surpassing high threshold, the load in certain territory just divided; When the load in certain territory during lower than lowest threshold, just it and its adjacent domains are merged, reduce the quantity in territory, finally between data processor, carry out domain scheduling, guarantee the load balancing between each data processor, thereby effectively eliminate the serious load imbalance that in extensive simulation application, group's migration brings, improve the utilance of large-scale cluster, accelerate the simulation program speed of service.

As shown in Figure 1, load-balancing method in the extensive Behavior modeling application of the present invention is to be applied in a kind of system (abbreviation runtime system) of support program operation, this system comprises task manager and data processor, wherein task manager is for initialization and the end of management data processor, data processor is for the treatment of the data of distributing to it, and comprise load equalizer of the present invention, data management system and tasks carrying device, load equalizer is for the load between balanced each data processor, data management system is distributed to the data of its affiliated data processor for management, the tasks carrying device is for carrying out the data that user program comes the deal with data manager to manage.

As shown in Figure 2, load-balancing method in the extensive Behavior modeling application of the present invention is to be applied in a kind of load equalizer, the data of this load equalizer for moving its place data processor data management system, to realize the load balancing between data processor, the method comprises the following steps:

(1) set up two Hash tables, one of them is for depositing all simulated object of piece, and another is for depositing all block messages in territory, and this Hash table is carried out to initialization, and this step comprises following sub-step:

(1-1) set up a piece table and a territory table.The piece table is for depositing all simulated object of piece, the corresponding data group of each list item, and comprise that four are listd: first lists the key assignments for memory block, lising for second is a chained list, the key assignments that is used for the adjacent block of memory block, the 3rd simulated object number of lising for memory block, lising for the 4th is a chained list, for all simulated object information of memory block, it is specifically as shown in table 1.

The piece key assignments	The adjacent block key assignments	The simulated object number	The simulated object information comprised
				?	?	?	?

?

?

?

?

Table l

The territory table is for depositing all block messages in territory, each list item is a corresponding data group also, and comprise that four are listd: first is listd for storing the key assignments in territory, list for storing the load in territory for second, list adjacent domains information for storing territory (comprising key assignments, place data processor number and the load of adjacent domains) for the 3rd, lising for the 4th is a chained list, the key assignments of all comprised for storing territory, and it is specifically as shown in table 2;

The territory key assignments	The territory load	Adjacent domains information	The piece key assignments comprised
				?	?	?	?
?	?	?	?

Table 2

The advantage of this sub-step is, by setting up Hash table, can be conducive to fast finding, insertion and the renewal in piece and territory.

(1-2) the whole virtual space of extensive Behavior modeling application is divided into to the piece of fixed size, the information of all is filled out in the write-in block table, and be averagely allocated to a plurality of territories by all; The quantity in territory and user-defined number of tasks are equal, and the quantity of piece is to be specified by the user;

(1-3) according to the load in all territories of following formula initialization:

$\left\{\begin{array}{c}\mathrm{Cost}\left(D\right)=\underset{P\∈D}{\mathrm{\Σ}}\mathrm{Cost}\left(P\right)\\ \mathrm{Cost}\left(P\right)=\left|P\right|\×\underset{Q\∈\mathrm{Neighbors}\left(P\right)}{\mathrm{\Σ}}\left|Q\right|,\end{array}\right.$

Wherein Cost (D) and the load of Cost (P) difference representative domain D and the load of the piece P in IncFlds D, and | P| and | Q | mean respectively the simulated object quantity that piece P and piece Q comprise, these two values can obtain by the query block table, wherein Neighbors (P) means the adjacent block set of piece P, and this value also can obtain by the query block table.

(1-4) notice tasks carrying device starts the execution of extensive Behavior modeling application;

The advantage of this step is, can know each take turns analog computation and complete after the position of the residing new piece of model objects.

(2) repeat increment calculates the load in each territory, until extensive Behavior modeling application end, with the territory D after the load of generation incremental computations; This step comprises following sub-step:

(2-1) judge whether that the processed complete simulated object of all new rounds is all processed, if yes then enter step (2-8), otherwise enter step (2-2);

(2-2) obtain the new location information (O of the simulated object O after processed in each territory _x, O _y, O _z);

(2-3) according to the new location information of simulated object and calculate the numbering K of this simulated object place piece according to following formula _new:

$\left\{\begin{array}{c}x=\mathrm{Ox}/S_x\\ y=\mathrm{Oy}/S_y\\ z=\mathrm{Oz}/S_z\\ \mathrm{Knew}=x+y\×\mathrm{Nx}+z\×\mathrm{Nx}\×\mathrm{Ny}\end{array}\right.$

N wherein _x, N _yand N _zthe extensive Behavior modeling application simulation space x in user-defined edge, the number of blocks of y and z axle, and S _x, S _yand S _zbe piece respectively along x, the size of y and z axle, each piece is S _x* S _y* S _zsize.Obvious S _x, S _yand S _zcan by by virtual space along x, the size of y and z axle is respectively divided by number of blocks N _x, N _yand N _zobtain;

(2-4) the block number K of gained is calculated in judgement _newwhether identical with the block number at simulated object O place after processed, judge whether simulated object O moves between piece.If different, enter step (2-5), otherwise return to step (2-1);

(2-5) by this simulated object from the old K that is numbered _oldpiece in take out, and this simulated object is inserted and is numbered K _newpiece in, and upgrade in the piece table and be numbered K _oldbe numbered K _newthe simulated object of piece count information and simulated object list information;

(2-6) judgement is numbered K _oldbe numbered K _newthe simulated object number change value Δ of piece | P| whether surpass threshold value T(wherein T be more than or equal to 1 positive integer, generally be set as T and be more than or equal to 50), if surpass, just enter step (2-7), otherwise return to step (2-1);

The advantage of this sub-step is the calculating that can avoid invalid, and this is because a small amount of simulated object number changes the impact of load imbalance degree few;

(2-7) be numbered K according to following formula incremental update _oldbe numbered K _newthe load in territory, piece place and the load of this territory adjacent domains, then return to step (2-1),

$\left\{\begin{array}{c}\mathrm{Cost}\left(D\right)=\mathrm{Cost}\left(D\right)+\underset{P\∈D}{\mathrm{\Σ}}\mathrm{\ΔCost}\left(P\right)\\ \mathrm{\ΔCost}\left(P\right)=\left|P\right|\×\mathrm{\ΔN}+\mathrm{\Δ}\left|P\right|\×N+\mathrm{\Δ}\left|P\right|\×\mathrm{\ΔN}\\ \mathrm{\ΔN}=\underset{Q\∈\mathrm{Neighbors}\left(P\right)}{\mathrm{\Σ}}\mathrm{\Δ}\left|Q\right|\end{array}\right.$

Wherein Δ means the difference to dependent variable, be the difference that Δ Cost (P) means Cost (P), Δ N means the difference of N, Δ | P| means the difference of P, Δ | Q| means the difference of Q, and the advantage of this sub-step is to avoid a large amount of loads to recalculate the expense of bringing;

(2-8) judge whether that all territories all are disposed, if enter step (5), otherwise enter step (2-9);

(2-9) whether the load that judges the territory D after all incremental computations loads is greater than T _max(T wherein _maxbe to be more than or equal to 1 positive integer, be traditionally arranged to be T _maxbe more than or equal to 4), if so, enter step (3), otherwise enter step (2-10);

(2-10) whether the load that judges the territory D after all incremental computations loads is less than T _min(T wherein _minspan is 0<T _min≤ 1, be traditionally arranged to be 0<T _min≤ 1/4), if so, enter step (4), otherwise return to step (2-8);

The advantage of this step (2) is the load in each territory of calculating of increment, need to when each takes turns analog computation, not recalculate the load in all territories, thereby can effectively accurately obtain the load in each territory.

(3) territory D is divided to generate territory set D _set, the territory set D after wherein dividing _setin the load of each territory D2 be T1(wherein T1 be more than or equal to 1 positive integer), and piece wherein is all interconnective.

The advantage of this step is only to need to divide the great territory of load, before and after making, twice division result is basically identical, thereby reduce the Data Migration amount of step (5), and can guarantee that all in same territory interconnect, thereby better utilize naturally occurring data locality in the Behavior modeling application.

As shown in Figure 3, this step specifically comprises following sub-step:

(3-1) edge of newly-built territory D2，Cong territory D takes out a load and is less than the piece P1 of T1, and is inserted into territory D2;

(3-2) obtain all set of blocks P that are connected with piece P1 in the D of territory _set;

(3-3) from set of blocks P _setmiddle taking-up piece P2;

(3-4) after decision block P2 inserts territory D2, whether the load of D2 is less than T1, if so, enters step (3-5), otherwise enters step (3-7);

(3-5) piece P2 is inserted to D2, and upgrade the load of D2;

(3-6) obtain set of blocks adjacent with P2 in D, and it is incorporated to set P _set, then return to step (3-3);

(3-7) D2 is inserted to resultant field set D _set;

(3-8) whether the load that judges D still is greater than T _max, if so, return to step (3-1), otherwise return to step (2-8);

(4) by territory D and its territory set D after division _setin adjacent domains merged, the load of the territory D after wherein merging is T1, and the piece in the D of territory is all interconnective.

The advantage of this step is only to need to merge the minimum territory of load, before and after making, twice division result is basically identical, thereby the Data Migration amount while reducing step (5), and can guarantee that all in same territory interconnect, thereby better utilize naturally occurring data locality in the Behavior modeling application.

As shown in Figure 4, this step comprises following sub-step:

(4-1) obtain first adjacent domains D1 of territory D;

(4-2) judge that whether D1Yu territory, territory D is at same data processor, and their load sum is less than threshold value T1, if so, enters step (4-3), otherwise enter step (4-4);

(4-3) territory D1 is merged in the D of territory, and the more load of neofield D, step (4-4) then entered;

(4-4) obtain the next adjacent domains D1 of D;

(4-5) judge that whether D1 is empty, if so, returns to step (2-8), otherwise enters step (4-2);

(5) carry out the domain scheduling between each data processor, realize the load balancing of data processor, wherein each territory of same data processor interconnects.

The advantage of this step is to guarantee that the territories of same data processor processes adjoin each other as far as possible, thereby better utilizes naturally occurring data locality in the Behavior modeling application.

This step specifically comprises following sub-step:

(5-1) load that judges current data processor whether surpass threshold value Tt(wherein Tt be user-defined number of tasks divided by the data processor number, its value is for being more than or equal to 8), if so, enter step (5-2), otherwise enter step (6);

(5-2) obtain a territory D of current data processor;

(5-3) obtain the data processor set W at all territory place adjacent with territory D;

(5-4) take out the data processor w1 of least-loaded in W;

(5-5) whether the load that judgement moves to territory D after data processor w1 is less than Tt, if so, enters step (5-6), otherwise enters step (5-7);

(5-6) territory D is moved to data processor w1 above, then enter step (5-1);

(5-7) obtain the data processor WL of least-loaded on all data processors, territory D is moved to this data processor WL above, and enter step (5-1);

(6) notice tasks carrying device starts the execution of an extensive Behavior modeling application new round;

(7) judge that whether extensive Behavior modeling program new round calculating finishes, and if so, enters step (8), otherwise returns to step (7);

(8) judge whether extensive Behavior modeling program has reached user-defined termination condition, if so, finishes the operation of load equalizer, otherwise return to step (2); Particularly, user-defined termination condition can be the wheel number repeatedly calculated.

Those skilled in the art will readily understand; the foregoing is only preferred embodiment of the present invention; not in order to limit the present invention, all any modifications of doing within the spirit and principles in the present invention, be equal to and replace and improvement etc., within all should being included in protection scope of the present invention.

Claims (6)

1. the load-balancing method in an extensive Behavior modeling application, to be applied in a kind of load equalizer, the data of this load equalizer for moving its place data processor data management system, to realize the load balancing between data processor, it is characterized in that, said method comprising the steps of:

(1) set up two Hash tables, one of them is for depositing all simulated object of piece, and another is for depositing all block messages in territory, and this Hash table is carried out to initialization;

(2) repeat increment calculates the load in each territory, until extensive Behavior modeling application end, with the territory D after the load of generation incremental computations; This step comprises following sub-step:

(2-1) judge whether that the processed complete simulated object of all new rounds is all processed, if yes then enter step (2-8), otherwise enter step (2-2);

(2-2) obtain the new location information (O of the simulated object O after processed in each territory _x, O _y, O _z);

(2-3) according to the new location information of simulated object and calculate the numbering K of this simulated object place piece according to following formula _new:

$\left\{\begin{array}{c}x=\mathrm{Ox}/S_x\\ y=\mathrm{Oy}/S_y\\ z=\mathrm{Oz}/S_z\\ \mathrm{Knew}=x+y\&times;\mathrm{Nx}+z\&times;\mathrm{Nx}\&times;\mathrm{Ny}\end{array}\right.$

N wherein _x, N _yand N _zthe extensive Behavior modeling application simulation space x in user-defined edge, the number of blocks of y and z axle, S _x, S _yand S _zbe piece respectively along x, the size of y and z axle;

(2-4) the block number K of gained is calculated in judgement _newwhether identical with the block number at simulated object O place after processed, if different, enter step (2-5), otherwise return to step (2-1);

(2-5) by this simulated object from the old K that is numbered _oldpiece in take out, and this simulated object is inserted and is numbered K _newpiece in, and upgrade in the piece table and be numbered K _oldbe numbered K _newthe simulated object of piece count information and simulated object list information;

(2-6) judgement is numbered K _oldbe numbered K _newthe simulated object number change value Δ of piece | whether P| surpasses threshold value T, if surpass, just enters step (2-7), otherwise returns to step (2-1);

(2-7) incremental update is numbered K _oldbe numbered K _newthe load in territory, piece place and the load of this territory adjacent domains, then return to step (2-1);

(2-8) judge whether that all territories all are disposed, if enter step (5), otherwise enter step (2-9);

(2-9) whether the load that judges the territory D after all incremental computations loads is greater than T _max, if so, enter step (3), otherwise enter step (2-10);

(2-10) whether the load that judges the territory D after all incremental computations loads is less than T _min, if so, enter step (4), otherwise return to step (2-8);

(3) territory D is divided to generate territory set D _set, the territory set D after wherein dividing _setin the load of each territory D2 be T1, and piece wherein is all interconnective; This step comprises following sub-step:

(3-1) edge of newly-built territory D2，Cong territory D takes out a load and is less than the piece P1 of T1, and is inserted into territory D2;

(3-2) obtain all set of blocks P that are connected with piece P1 in the D of territory _set;

(3-3) from set of blocks P _setmiddle taking-up piece P2;

(3-4) after decision block P2 inserts territory D2, whether the load of D2 is less than T1, if so, enters step (3-5), otherwise enters step (3-7);

(3-5) piece P2 is inserted to D2, and upgrade the load of D2;

(3-6) obtain set of blocks adjacent with P2 in D, and it is incorporated to set P _set, then return to step (3-3);

(3-7) D2 is inserted to resultant field set D _set;

(3-8) whether the load that judges D still is greater than T _max, if so, return to step (3-1), otherwise return to step (2-8);

(4) by territory D and its territory set D after division _setin adjacent domains merged, the load of the territory D after wherein merging is T1, and the piece in the D of territory is all interconnective;

(5) carry out the domain scheduling between each data processor, realize the load balancing of data processor, wherein each territory of same data processor interconnects.

(6) notice tasks carrying device starts the execution of an extensive Behavior modeling application new round;

(7) judge that whether extensive Behavior modeling program new round calculating finishes, and if so, enters step (8), otherwise returns to step (7);

(8) judge whether extensive Behavior modeling program has reached user-defined termination condition, if so, finishes the operation of load equalizer, otherwise return to step (2).

2. load-balancing method according to claim 1, is characterized in that, step (1) comprises following sub-step:

(1-1) set up a piece table and a territory table, wherein the piece table is for depositing all simulated object of piece, the corresponding data group of each list item, and comprise that four are listd: first lists the key assignments for memory block, lising for second is a chained list, the key assignments that is used for the adjacent block of memory block, the 3rd simulated object number of lising for memory block, lising for the 4th is a chained list, all simulated object information for memory block, the territory table is for depositing all block messages in territory, each list item is a corresponding data group also, and comprise that four are listd: first is listd for storing the key assignments in territory, list for storing the load in territory for second, list for storing the adjacent domains information in territory for the 3rd, lising for the 4th is a chained list, for storing the key assignments of all that territory comprises,

(1-2) the whole virtual space of extensive Behavior modeling application is divided into to the piece of fixed size, the information of all is filled out in the write-in block table, and be averagely allocated to a plurality of territories by all;

(1-3) load in all territories of initialization:

(1-4) notice tasks carrying device starts the execution of extensive Behavior modeling application.

3. load-balancing method according to claim 2, is characterized in that, step (1-3) is to adopt following formula:

$\left\{\begin{array}{c}\mathrm{Cost}\left(D\right)=\underset{P\&Element;D}{\mathrm{\&Sigma;}}\mathrm{Cost}\left(P\right)\\ \mathrm{Cost}\left(P\right)=\left|P\right|\&times;\underset{Q\&Element;\mathrm{Neighbors}\left(P\right)}{\mathrm{\&Sigma;}}\left|Q\right|,\end{array}\right.$

Cost(D wherein) and the Cost(P) load of representative domain D and the load of the piece P in IncFlds D respectively, and | P| and | Q| means respectively the simulated object quantity that piece P and piece Q comprise, and Neighbors(P) means the adjacent block set of piece P.

4. load-balancing method according to claim 3, is characterized in that, step (2-7) is to adopt following formula:

$\left\{\begin{array}{c}\mathrm{Cost}\left(D\right)=\mathrm{Cost}\left(D\right)+\underset{P\&Element;D}{\mathrm{\&Sigma;}}\mathrm{\&Delta;Cost}\left(P\right)\\ \mathrm{\&Delta;Cost}\left(P\right)=\left|P\right|\&times;\mathrm{\&Delta;N}+\mathrm{\&Delta;}\left|P\right|\&times;N+\mathrm{\&Delta;}\left|P\right|\&times;\mathrm{\&Delta;N}\\ \mathrm{\&Delta;N}=\underset{Q\&Element;\mathrm{Neighbors}\left(P\right)}{\mathrm{\&Sigma;}}\mathrm{\&Delta;}\left|Q\right|,\end{array}\right.$

Wherein Δ means the difference to dependent variable.

5. load-balancing method according to claim 1, is characterized in that, step (4) comprises following sub-step:

(4-1) obtain first adjacent domains D1 of territory D;

(4-2) judge that whether D1Yu territory, territory D is at same data processor, and their load sum is less than threshold value T1, if so, enters step (4-3), otherwise enter step (4-4);

(4-3) territory D1 is merged in the D of territory, and the more load of neofield D, step (4-4) then entered;

(4-4) obtain the next adjacent domains D1 of D;

(4-5) judge that whether D1 is empty, if so, returns to step (2-8), otherwise enters step (4-2).

6. load-balancing method according to claim 1, is characterized in that, step (5) comprises following sub-step:

(5-1) whether the load that judges current data processor, over threshold value Tt, if so, enters step (5-2), otherwise enters step (6);

(5-2) obtain a territory D of current data processor;

(5-3) obtain the data processor set W at all territory place adjacent with territory D;

(5-4) take out the data processor w1 of least-loaded in W;

(5-5) whether the load that judgement moves to territory D after data processor w1 is less than Tt, if so, enters step (5-6), otherwise enters step (5-7);

(5-6) territory D is moved to data processor w1 above, then enter step (5-1);

(5-7) obtain the data processor WL of least-loaded on all data processors, territory D is moved to this data processor WL above, and enter step (5-1).

Images