CN102508708A - Heterogeneous multi-core energy-saving task schedule method based on improved genetic algorithm - Google Patents

Abstract

A heterogeneous multi-core energy-saving task schedule method based on improved genetic algorithm comprises the improved genetic algorithm used for determining task priority and energy-saving schedule algorithm based on zooming priority. The flow of the heterogeneous multi-core energy-saving task schedule method includes (1), initializing population information; (2), entering a loop body and determining the task priority by the aid of genetic algorithm; (3), determining schedule sequence of tasks on a processor according to a task DAG (directed acyclic graph) and division strategies; (4), realizing dynamic voltage zooming on the basis of feasible task schedule according to relation of saving energy of the tasks and prolonging time; (5), calculating fitness of a current population and sorting the current population; and (6), updating the population by the aid of the improved genetic algorithm, determining new task priority, quitting if termination conditions are met, and continuing iteration if the termination conditions are not met.

Description

Based on the energy-conservation method for scheduling task of heterogeneous polynuclear that improves genetic algorithm

Technical field

The present invention is mainly concerned with the design field of embedded system, refers in particular to a kind of based on the energy-conservation method for scheduling task of heterogeneous polynuclear that improves genetic algorithm.

Background technology

For solving the power problems in the Embedded System Design, the dynamic voltage scaling technology of rising in recent years has caused the extensive concern of industry.At present, many popular low power processors are all supported DVS (dynamic voltage scaling technology) technology, and the DVS technology promptly is to reduce system power dissipation through the voltage of real time altering processor unit and frequency.Yet the system frequency reduction will cause task execution time elongated, possibly influence the real-time of system.Thereby, in the voltage scaling process, need to select suitable dispatching method, under the situation that does not influence system performance, power consumption is optimized.

Under other heterogeneous multiprocessor system of discrete voltage level associated task is carried out energy-saving distribution and be proved to be the problem that belongs to very difficult already having, need to adopt heuritic approach that it is found the solution usually.

In the prior art; What have is expressed as the voltage scaling problem integer programming problem; Wherein representational is the energy perception algorithm of dispatching many speed periodic duty and aperiodic task simultaneously, but it is too big to be based on the algorithm complex of integer programming, is difficult in the reality and is used widely.

Some researcher combines random algorithm and energy gradient technology to solve the distribution and the task scheduling problem of relaxation time sheet simultaneously; The energy-conservation otherness that algorithm but average or Random assignment sheet slack time has been ignored different task can't realize optimum energy-saving effect.

The also energy gradient of with good grounds task and execution time calculating priority level; Under priority instructs, carry out then random schedule and based on the tabulation dispatching algorithm of priority; But the dispatching algorithm based on priority has very big limitation on solution space merely, is difficult to obtain system optimal and separates.

Other have with traditional genetic algorithm with the tabulation scheduling strategy combine; The execution sequence that heredity tabulation dispatching algorithm sets the tasks is proposed; Accomplish the energy-saving distribution under the heterogeneous polynuclear environment in conjunction with relevant dynamic voltage scaling algorithm; Though but the dispatching algorithm that is based on traditional genetic algorithm has wide solution space, the local optimal searching function is not strong, thereby influences energy-saving effect.

Summary of the invention

The technical matters that the present invention will solve just is: to the technical matters that prior art exists, the present invention provides that a kind of principle is simple, working time short, have preferable energy-saving effect based on the energy-conservation method for scheduling task of heterogeneous polynuclear that improves genetic algorithm.

For solving the problems of the technologies described above, the present invention adopts following technical scheme:

A kind of based on the energy-conservation method for scheduling task of heterogeneous polynuclear that improves genetic algorithm; It is characterized in that; By being used for setting the tasks the improvement genetic algorithm of priority and form based on the energy-saving distribution algorithm of convergent-divergent priority, its flow process is: the population information initializing is carried out in (1); (2) get into loop body through the genetic algorithm priority that sets the tasks; (3) according to task DAG figure and partition strategy, dispatching sequence on processor sets the tasks; (4) save energy and the relation between time expand according to task, at the feasible enterprising action attitude voltage scaling in task scheduling basis; (5) calculate current colony fitness and ordering; (6) adopt improved genetic algorithm that population is upgraded, confirm new task priority,, otherwise continue iteration if satisfy end condition then withdraw from.

As further improvement of the present invention:

In the said step (1), convert task priority into the chromosome of tissue in certain sequence, wherein chromosomal coding adopts one dimension bit string form; Code length is that task node is counted n in the task image, the dispatching priority of the corresponding task of each gene representation, and span is [0; N-1]; The low more corresponding priority of numerical value is high more, for the identical task of priority, then sorts according to mission number.

The flow process of said step (6) is:

(6.1) colony is carried out selection operation, from current colony, select individuality, intersect and mutation operation as parent; Select the operator execution in step following:

The individual number Nsel that (6.1.1) selects according to calculation of parameter needs such as number of groups in the genetic algorithm;

(6.1.2) according to the ranking results of front ordering, select to be positioned at the anterior Nsel/2 individuals of formation;

(6.1.3) adopt random algorithm, in formation, evenly produce the Nsel/2 individuals in all the other individualities;

(6.1.4) with above-mentioned (6.1.2) and the common Nsel individuals that (6.1.3) produces intersect as selection result and make a variation the selection course end;

(6.2) carry out interlace operation: adopt 2 crossover operators;

(6.3) mutation operation: mutation operator is that certain gene in the individuality is carried out stochastic transformation in valid value range, also is certain task executions priority of randomly changing;

(6.4) colony upgrades, and execution in step is following:

(6.4.1) the scheduling energy consumption Echild of the newly-generated individuality of calculating;

(6.4.2) Echild and its former generation's scheduling energy consumption Eparent relatively divides two kinds of situation to consider: if Echild less than Eparent, then selects the chromosome of new generation; If Echild more than or equal to Eparent, then accepts new chromosome with probability P receive, the computing method following formula of Preceive:

$\left\{\begin{array}{c}{P}_{\mathrm{receive}}=e^{-(E_{\mathrm{child}}-E_{\mathrm{parent}})/T}\\ T=\frac{1}{P}\underset{i=1}{\overset{P}{\mathrm{\Σ}}}{(E_i-\stackrel{\‾}{E})}^2\end{array}\right.$

Wherein, E is a natural constant; Ei is the pairing energy consumption of individual i;

on average carries out energy consumption for colony, and P is a population size;

(6.4.3) go to upgrade colony with the new chromosome of selecting;

If (6.4.4) also have new chromosome to be untreated, then forward (6.4.1) to, upgrade otherwise finish colony.

The flow process of said energy-saving distribution algorithm based on convergent-divergent priority is:

(4.1) carry out algorithm initialization;

(4.2) all tasks are composed with ceiling voltage;

(4.3) adopt priority chain list scheduling algorithm to carry out task scheduling, algorithm is according to input informations such as duty mapping, voltage level and task priorities, and the output of determinacy ground task scheduling strategy, execution time and energy consumption can not be dispatched and then returned an invalid value;

(4.4) other scheduling result of ceiling voltage level is judged, do not had scale space, return the scheduling energy if the execution time, is explained this scheduling more than or equal to closing time.

(4.5) carry out voltage convergent-divergent process repeatedly, wherein to each can convergent-divergent task operate as follows:

Reduce the laggard row major level of a voltage level chain list scheduling, if satisfy the off period, the convergent-divergent priority of calculation task Ti under current voltage then, priority is following formula:

$\mathrm{ZoomPriority}(i,k)=K_{\mathrm{coe}}\·\frac{E(T_i,P_j,V_k)-E(T_i,P_j,V_k-1)}{t(T_i,P_j,V_k-1)-t(T_i,P_j,V_k)}$

In the formula, and ZoomPriority (i, k): task T _iConvergent-divergent priority under current voltage.K _CoeBe scale-up factor, E (T _i, P _j, V _k): task T _iIn processor P _jIn with electric pressure V _kThe energy consumption of operation.T (T _i, P _j, V _k): task T _iIn processor P _jIn with electric pressure V _kThe time of operation;

(4.6) maximum so far scalable priority of record and corresponding scheduling result thereof;

(4.7) feasible optimum convergent-divergent task is reduced a voltage level, and if upgrade the current optimal energy consumption and do not find scalable task then return the current optimal power consumption values, algorithm withdraws from.

Compared with prior art; The invention has the advantages that: the present invention is through improving traditional genetic algorithm; Avoided it to be absorbed in the shortcoming of local optimum easily; Can satisfy the energy-saving distribution that uses under the solution space based on convergent-divergent priority, adopt the dynamic voltage scaling technology under the heterogeneous polynuclear environment, to realize energy-conservation task scheduling.The present invention influences the multifarious selection operator of colony and colony's update mechanism at first to the not good defective of traditional genetic algorithm local search ability through improving, and expands the solution space of genetic algorithm, and dispatching priority sets the tasks; Then, the individuality to genetic algorithm is confirmed has designed a kind of energy-saving distribution algorithm based on voltage scaling priority, carries out voltage-regulation through the task of selecting repeatedly to have best zooming effect, optimizes each individual scheduling energy consumption; At last, select optimum scheduling scheme, realize the energy-conservation task scheduling under the heterogeneous polynuclear environment through the genetic iteration operation.Experimental result shows, the synthesis energy saving rate and working times two the aspect factor, the present invention has obviously superior overall performance.

Description of drawings

Fig. 1 is a kind of synoptic diagram of DAG task image instance;

Fig. 2 is a schematic flow sheet of the present invention;

Fig. 3 is based on the schematic flow sheet of the energy-saving distribution algorithm of convergent-divergent priority among the present invention.

Embodiment

Below will combine Figure of description and specific embodiment that the present invention is explained further details.

Under the heterogeneous polynuclear environment; Task image to dependence before and after having carries out the energy consumption optimal scheduling; This problem can be described below: Task Distribution is arrived rational processor; And confirm the working voltage of this processor, and satisfy under the prerequisite of off period of all tasks, realize entire system energy consumption optimal target.Generally speaking, the scheduling strategy of on heterogeneous multi-nucleus processor, seeking energetic optimum can be divided into three phases, is respectively task division, task scheduling and dynamic voltage scaling.Task division is to each processor with Task Distribution; Task scheduling on the basis of task division, the execution sequence that sets the tasks; Dynamic voltage scaling is the execution voltage that sets the tasks.

Earlier to involved in the present invention to energy model, system model and task model describe.

For example, for CMOS (complementary metal oxide semiconductor (CMOS)) circuit, dynamic power consumption is the main source that system capacity consumes.For same processing unit, P _Dyn(dynamic power consumption) and V _DdRelation between (supply voltage of system) and f (operating frequency) three is as shown in the formula (1) and formula (2).

$f=K_{\mathrm{ic}}\·\frac{{(V_{\mathrm{dd}}-V_t)}^2}{V_{\mathrm{dd}}}---\left(1\right)$

$P_{\mathrm{dyn}}=C_{\mathrm{ef}}\&CenterDot;{\mathrm{<figure-callout\; id="2"\; label="V\; dd"\; filenames="HDA0000114535910000011.png"\; state="\{\{state\}\}">V</figure-callout>}}_{\mathrm{dd}}^{}\&CenterDot;f---\left(2\right)$

Wherein, C _EfThe efficient loading electric capacity of indication circuit, K _IcRepresent the constant relevant, V with circuit _tThe threshold voltage of the normal operation of expression system.What formula (1) was represented is the relation between supply voltage and the processor frequencies, and what formula (2) disclosed is the relation between dynamic power consumption, supply voltage, the processor frequencies three.

In addition, according to the relation between task execution time and the system energy consumption, obtain formula (3) and (4):

$t=\frac{N_c}{f}---\left(3\right)$

E _dyn＝P _dyn·t (4)

Wherein, N _cThe expression task is carried out required processor operations number of times, confirms that by task attribute and corresponding processing device structure thereof t is a task execution time, E _DynThe dynamic energy consumption of expression system according to formula (2), (3), (4), obtains formula (5):

$E_{\mathrm{dyn}}=C_{\mathrm{ef}}\&CenterDot;{\mathrm{<figure-callout\; id="2"\; label="V\; dd"\; filenames="HDA0000114535910000011.png"\; state="\{\{state\}\}">V</figure-callout>}}_{\mathrm{dd}}^{}\&CenterDot;N_c---\left(5\right)$

Can find out from formula (5), if the processor that task is shone upon is confirmed N then _cBe a constant, thereby the processor dynamic power consumption become a square proportional relation with supply voltage.If want to reduce power consumption of processing unit, then can realize through reducing system voltage.But the supply voltage of system can not reduce arbitrarily, and it can not be lower than the normal working voltage threshold values V of system _tCan know by formula (1), satisfy V _Dd＞V _tPrerequisite under, frequency f is about supply voltage V _DdIncreasing function, when voltage reduced, the frequency of processor also decreased, and so just possibly cause the task off period to miss, thus the definite voltage scaling that how reasonably to carry out of needs.

This heterogeneous multiprocessor system by a series of possess the dynamic voltage scaling function but processing unit that performance there are differences (Processing Element PE) forms, and is expressed as PE={PE1, PE2 ..., PEm}, each processing unit PE _j(j=1,2 ... m) have some discrete voltage mode V (j, k), k=1 wherein, 2 ... N (j), N (j) expression processing unit PE _jThe different electric die pressing type that has.PE _jVoltage mode V (j, power consumption under k) and frequency use respectively P (j, k) and f (j k) representes.Adopt bus mode to communicate between the different processing units, the power consumption of bus is P _bThereby total dynamic energy consumption Etotal can be by representing with formula (6) in the task implementation:

$E_{\mathrm{total}}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{N_j}{\mathrm{\&Sigma;}}}(\mathrm{\&theta;}(i,j,k)\&times;P(i,j,k)\&times;t_{\mathrm{exe}}(i,j,k))+P_b\&CenterDot;t_c---\left(6\right)$

Wherein, n, m represent number of tasks and processor unit number respectively, and (i, j k) represent task T to P _iAt processing unit PE _jGo up (j, the dynamic power consumption when k) carrying out, t with voltage level V _Exe(i, j k) are the corresponding execution time, and θ (i, j k) are a choice function, and it has 0,1 two value, have only task T _iBe divided into processing unit PE _jGo up and with voltage mode V (j, when k) carrying out, (value k) just is 1 to θ, otherwise is 0 for i, j; t _cBe the total call duration time between the task under the current division.

For the task-set in the dispatching algorithm, (Directed Acyclic Graph DAG) representes to adopt directed acyclic graph usually.Fig. 1 is a simple DAG task image, and each node is represented a task, and the communication between the task is represented on the limit between the node, numeral be communication cost.Have dependence between the task, for example, task T6 is only complete at its all father node T4 and T5, and result of calculation could be carried out after sending.

Here adopt the DAG task image, task-set with doublet G (T, C) expression, T={T1 wherein, T2 ..., Tn} representes n set of tasks to be carried out, C is the set of communication limit, C (i, j) expression task T _iTo T _jBetween communication overhead.Here the task model that adopts also has following character:

(1) periodic duty model, task-set arrived according to certain cycle, and each task-set must be accomplished in off period tdeadline,

(2) task can not be seized, and begins on the unit to carry out in case task is managed somewhere, finishes just to discharge this processing unit until it.

(3) have before and after the task of dependence when being assigned to different processing units, the communication overhead between then must the consideration task; Dependence task on the same processor is not considered call duration time.

The present invention is under the prerequisite of dividing that sets the tasks, proposition a kind of based on the energy-conservation method for scheduling task of heterogeneous polynuclear that improves genetic algorithm, and it is by being used for setting the tasks the improvement genetic algorithm of priority and form based on the energy-saving distribution algorithm of convergent-divergent priority.As shown in Figure 2, method of the present invention is at first carried out the population information initializing, gets into loop body then.Earlier through genetic algorithm is improved, priority sets the tasks in loop body; Again according to task DAG figure and partition strategy, dispatching sequence on processor sets the tasks; Save energy and the relation between time expand according to task at last, at the feasible enterprising action attitude voltage scaling in task scheduling basis; Next, calculate current colony fitness and ordering; Adopt improved genetic algorithm that population is upgraded, confirm new task priority,, otherwise continue iteration if satisfy end condition then withdraw from.

The traditional genetic algorithm is absorbed in locally optimal solution easily when solving like challenges such as task schedulings.When the key that addresses this problem is to keep algorithm the convergence speed, also to keep the diversity of colony.The present invention improves through selection operator and the population update strategy to traditional genetic algorithm; Adding is to the selection probability of the not good enough individuality of temporary transient performance; Increase the diversity of population; Improve the overall situation and the local optimal searching ability of genetic algorithm, improved genetic algorithm is used for carrying out energy-conservation task scheduling on the heterogeneous multiprocessor system, thereby realizes energy-conservation effect.

Detailed process of the present invention is:

1. colony is carried out initialization: convert task priority into the chromosome of tissue in certain sequence.Wherein chromosomal coding adopts one dimension bit string form; Code length is that task node is counted n in the task image, the dispatching priority of the corresponding task of each gene representation, and span is [0; N-1]; The low more corresponding priority of numerical value is high more, for the identical task of priority, then sorts according to mission number.DAG figure as shown in Figure 1, it comprises 9 tasks, and a kind of possible chromosome coding is 342587164, and the priority of expression task 1 is 3, and the priority of task 2 is 4, by that analogy.

2. energy-saving distribution and ordering: at first progressively the energy-saving distribution algorithm based on convergent-divergent priority among the employing of the individuality in colony the present invention is carried out energy-saving distribution; The input informations such as task priority that this algorithm is confirmed according to genetic algorithm return individual schedule strategy and corresponding energy consumption, time etc.; Then the scheduling energy consumption of individuality is carried out ascending sort, what energy consumption was minimum is positioned at formation foremost.

3. genetic iteration:

3.1 colony is carried out selection operation, from current colony, select individuality, intersect and mutation operation as parent.

The present invention will preferentially select operator and select operator to combine at random through taking all factors into consideration good property of population and diversity, improve selecting operator.Through to selecting operator to improve, under the prerequisite that does not increase algorithm complex, can guarantee the quality and the diversity of colony simultaneously, reduce the possibility that it is absorbed in local optimum.Improve and select the operator execution in step following:

(1) the individual number Nsel that selects according to calculation of parameter needs such as number of groups in the genetic algorithm;

(2), select to be positioned at the anterior Nsel/2 individuals of formation according to the ranking results of front ordering;

(3) adopt random algorithm, in formation, evenly produce the Nsel/2 individuals in all the other individualities;

(4) the common Nsel individuals that (2), (3) step is produced is intersected as selection result and is made a variation, and selection course finishes.

3.2 carry out interlace operation: the present invention adopts 2 crossover operators, and this operator is simple to operate for the chromosome of one-dimension array coding, and can expand solution space preferably.

3.3 mutation operation: mutation operator is that certain gene in the individuality is carried out stochastic transformation in valid value range, also is certain task executions priority of randomly changing, thus the diversity of expansion solution space.

3.4. colony upgrades: for genetic algorithm, multifarious colony is beneficial to and forms wide solution space, prevents to be absorbed in local optimum, and then obtains globally optimal solution with more great probability.Colony's update mechanism of traditional genetic algorithm is directly upgraded colony with more excellent individuality in filial generation and the parent, and the energy-conservation task scheduling of this strategy often causes precocious phenomenon.To this specific background of the energy-conservation task scheduling of heterogeneous polynuclear, the present invention adopts the Metropolis acceptance criterion that colony's update mechanism in the traditional genetic algorithm is improved.

The Metropolis acceptance criterion is the importance sampling method that people such as Metropolis proposed in nineteen fifty-three, and it both can tend to the direction search of objective optimization, can accept bad separating according to certain probability again.This acceptance criteria is generally used for the simulated annealing process, and can it be extrapolated in the middle of the energy-saving distribution, and improved colony update mechanism is following:

(1) the scheduling energy consumption Echild of the newly-generated individuality of calculating;

(2) Echild and its former generation's scheduling energy consumption Eparent relatively divides two kinds of situation to consider: if Echild less than Eparent, then selects the chromosome of new generation; If Echild more than or equal to Eparent, then accepts new chromosome with probability P receive, the computing method of Preceive are suc as formula (7).

(3) go to upgrade colony with the new chromosome of selecting;

(4) if also have new chromosome to be untreated, then forward (1) to, upgrade otherwise finish colony.

Formula (7) is confirmed new chromosomal probability of acceptance Preceive according to Metropolis acceptance criterion and task energy, and computing method are as follows:

$\left\{\begin{array}{c}{P}_{\mathrm{receive}}=e^{-(E_{\mathrm{child}}-E_{\mathrm{parent}})/T}\\ T=\frac{1}{P}\underset{i=1}{\overset{P}{\mathrm{\&Sigma;}}}{(E_i-\stackrel{\&OverBar;}{E})}^2\end{array}\right.---\left(7\right)$

Wherein, E is a natural constant; Ei is the pairing energy consumption of individual i;

on average carries out energy consumption for colony, and P is a population size.

4. the validity of this iteration is judged; If this iteration and last iteration least energy consumption ratio less than significant indexes, explain that then filial generation has had remarkable improvement, invalid iterations uIter is put 0; And write down this iteration optimum solution, otherwise uIter is added 1.When reaching maximum iteration time ITERMAX or invalid iterations above UMAXITER, program withdraws from.

In the present embodiment; Based on the energy-saving distribution algorithm of convergent-divergent priority is inputs such as the task priority confirmed according to genetic algorithm and DAG task image, processor structure information, duty mapping; Carry out the voltage scaling energy-saving distribution, its output result be information such as the system's execution time, energy consumption under energy-saving distribution strategy and this strategy.

Before providing arthmetic statement, the Several Parameters of using in the definition algorithm earlier, as follows.

Referring to Fig. 3, in the present embodiment, be based on the flow process of the energy-saving distribution algorithm of convergent-divergent priority:

2.1 carry out algorithm initialization.

2.2 all tasks are composed with ceiling voltage.

2.3 (Priority List Schedule Algorithm PLSA) carries out task scheduling, and this algorithm is classical polynomial time task scheduling algorithm to adopt priority chain list scheduling algorithm.Algorithm is according to input informations such as duty mapping, voltage level and task priorities, and the output of determinacy ground task scheduling strategy, execution time and energy consumption can not be dispatched and then returned an invalid value.

2.4 other scheduling result of ceiling voltage level is judged, if, explaining this scheduling more than or equal to closing time, the execution time do not have scale space, return the scheduling energy.

2.5 carry out voltage convergent-divergent process repeatedly.Wherein to each can convergent-divergent task operate as follows:

Reduce the laggard row major level of a voltage level chain list scheduling, if satisfy the off period, the convergent-divergent priority of calculation task Ti under current voltage then.Priority is calculated suc as formula (8).

$\mathrm{ZoomPriority}(i,k)=K_{\mathrm{coe}}\&CenterDot;\frac{E(T_i,P_j,V_k)-E(T_i,P_j,V_k-1)}{t(T_i,P_j,V_k-1)-t(T_i,P_j,V_k)}---\left(8\right)$

In the formula (8), and ZoomPriority (i, k): task T _iConvergent-divergent priority under current voltage.K _CoeBe scale-up factor, E (T _i, P _j, V _k): task T _iIn processor P _jIn with electric pressure V _kThe energy consumption of operation.T (T _i, P _j, V _k): task T _iIn processor P _jIn with electric pressure V _kThe time of operation.

The fraction molecule is that voltage reduces the energy that rank is practiced thrift, and denominator is the task run time that corresponding operating increased.

2.6 write down maximum so far scalable priority and corresponding scheduling result thereof.

2.7 feasible optimum convergent-divergent task is reduced a voltage level, and if upgrade the current optimal energy consumption and do not find scalable task then return the current optimal power consumption values, algorithm withdraws from.

All find out a feasible OPTIMAL TASK based on the each loop iteration of energy-saving distribution algorithm of convergent-divergent priority and carry out convergent-divergent, till can't proceeding convergent-divergent.Through iterating, individual energy consumption is optimized.

Below only be preferred implementation of the present invention, protection scope of the present invention also not only is confined to the foregoing description, and all technical schemes that belongs under the thinking of the present invention all belong to protection scope of the present invention.Should be pointed out that for those skilled in the art some improvement and retouching not breaking away under the principle of the invention prerequisite should be regarded as protection scope of the present invention.

Claims (4)

1. one kind based on the energy-conservation method for scheduling task of heterogeneous polynuclear that improves genetic algorithm; It is characterized in that; By being used for setting the tasks the improvement genetic algorithm of priority and form based on the energy-saving distribution algorithm of convergent-divergent priority, its flow process is: the population information initializing is carried out in (1); (2) get into loop body through the genetic algorithm priority that sets the tasks; (3) according to task DAG figure and partition strategy, dispatching sequence on processor sets the tasks; (4) save energy and the relation between time expand according to task, at the feasible enterprising action attitude voltage scaling in task scheduling basis; (5) calculate current colony fitness and ordering; (6) adopt improved genetic algorithm that population is upgraded, confirm new task priority,, otherwise continue iteration if satisfy end condition then withdraw from.

2. according to claim 1 based on the energy-conservation method for scheduling task of heterogeneous polynuclear that improves genetic algorithm, it is characterized in that: in the said step (1), convert task priority into the chromosome of tissue in certain sequence; Wherein chromosomal coding adopts one dimension bit string form; Code length is that task node is counted n in the task image, the dispatching priority of the corresponding task of each gene representation, and span is [0; N-1]; The low more corresponding priority of numerical value is high more, for the identical task of priority, then sorts according to mission number.

3. according to claim 1 based on the energy-conservation method for scheduling task of heterogeneous polynuclear that improves genetic algorithm, it is characterized in that the flow process of said step (6) is:

(6.1) colony is carried out selection operation, from current colony, select individuality, intersect and mutation operation as parent; Select the operator execution in step following:

The individual number Nsel that (6.1.1) selects according to calculation of parameter needs such as number of groups in the genetic algorithm;

(6.1.2) according to the ranking results of front ordering, select to be positioned at the anterior Nsel/2 individuals of formation;

(6.1.3) adopt random algorithm, in formation, evenly produce the Nsel/2 individuals in all the other individualities;

(6.1.4) with above-mentioned (6.1.2) and the common Nsel individuals that (6.1.3) produces intersect as selection result and make a variation the selection course end;

(6.2) carry out interlace operation: adopt 2 crossover operators;

(6.3) mutation operation: mutation operator is that certain gene in the individuality is carried out stochastic transformation in valid value range, also is certain task executions priority of randomly changing;

(6.4) colony upgrades, and execution in step is following:

(6.4.1) the scheduling energy consumption Echild of the newly-generated individuality of calculating;

(6.42) Echild and its former generation's scheduling energy consumption Eparent relatively divides two kinds of situation to consider: if Echild less than Eparent, then selects the chromosome of new generation; If Echild more than or equal to Eparent, then accepts new chromosome with probability P receive, the computing method following formula of Preceive:

$\left\{\begin{array}{c}{P}_{\mathrm{receive}}=e^{-(E_{\mathrm{child}}-E_{\mathrm{parent}})/T}\\ T=\frac{1}{P}\underset{i=1}{\overset{P}{\mathrm{\&Sigma;}}}{(E_i-\stackrel{\&OverBar;}{E})}^2\end{array}\right.$

Wherein, E is a natural constant; Ei is the pairing energy consumption of individual i;

on average carries out energy consumption for colony, and P is a population size;

(6.4.3) go to upgrade colony with the new chromosome of selecting;

If (6.4.4) also have new chromosome to be untreated, then forward (6.4.1) to, upgrade otherwise finish colony.

4. describedly it is characterized in that according to claim 1 or 2 or 3 that the flow process of said energy-saving distribution algorithm based on convergent-divergent priority is based on the energy-conservation method for scheduling task of heterogeneous polynuclear that improve genetic algorithm:

(4.1) carry out algorithm initialization;

(42) all tasks are composed with ceiling voltage;

(4.3) adopt priority chain list scheduling algorithm to carry out task scheduling, algorithm is according to input informations such as duty mapping, voltage level and task priorities, and the output of determinacy ground task scheduling strategy, execution time and energy consumption can not be dispatched and then returned an invalid value;

(4.4) other scheduling result of ceiling voltage level is judged, do not had scale space, return the scheduling energy if the execution time, is explained this scheduling more than or equal to closing time;

(4.5) carry out voltage convergent-divergent process repeatedly, wherein to each can convergent-divergent task operate as follows:

Reduce the laggard row major level of a voltage level chain list scheduling, if satisfy the off period, the convergent-divergent priority of calculation task Ti under current voltage then, priority is following formula:

$\mathrm{ZoomPriority}(i,k)=K_{\mathrm{coe}}\&CenterDot;\frac{E(T_i,P_j,V_k)-E(T_i,P_j,V_k-1)}{t(T_i,P_j,V_k-1)-t(T_i,P_j,V_k)}$

In the formula, and ZoomPriority (i, k): task T _iConvergent-divergent priority under current voltage.K _CoeBe scale-up factor, E (T _i, P _j, V _k): task T _iIn processor P _jIn with electric pressure V _kThe energy consumption of operation.T (T _i, P _j, V _k): task T _iIn processor P _jIn with electric pressure V _kThe time of operation;

(4.6) maximum so far scalable priority of record and corresponding scheduling result thereof;

(4.7) feasible optimum convergent-divergent task is reduced a voltage level, and upgrade the current optimal energy consumption, if do not find scalable task then return the current optimal power consumption values, algorithm withdraws from.

Images