CN102360246A - Self-adaptive threshold-based energy-saving scheduling method in heterogeneous distributed system - Google Patents

Abstract

The invention relates to a self-adaptive threshold-based energy-saving scheduling method in an environment of supporting a heterogeneous distributed system, belonging to the technical field of parallel task scheduling of the heterogeneous distributed system. The method provided by the invention specifically comprises the steps of: reading a parallel task directed acyclic graph (DAG) file; obtaining an initial task scheduling sequence; obtaining an optimal threshold: dynamically obtaining an optimal threshold according to parallel task and system environment; grouping the tasks: controlling task duplication by using the optimal threshold, namely, selectively duplicating optimal precursors of the tasks to balance system performance and energy consumption to obtain approximate optimal grouping; mapping the tasks: scheduling all groups into a processor, wherein the processor is not distributed and has minimum energy consumption; and regulating the voltage of the processor: dynamically regulating the voltage of the processor by using a task idle time to reduce the energy consumption of the processor. According to the invention, requirements on system performance and energy consumption are comprehensively considered. The optical threshold in the method can self-adapt the parallel tasks and the system environment, and is used for controlling the task duplication for balancing the system performance and the energy consumption, so that the energy consumption is reduced as far as possible under the premise that the final scheduling result meets the requirement of the system performance.

Description

In a kind of heterogeneous distributed system based on the energy-saving scheduling method of adaptive threshold

Technical field

The present invention relates to a kind of computer system energy-saving scheduling method, particularly relate to a kind of energy-saving scheduling method of supporting in the heterogeneous distributed system based on adaptive threshold.

Background technology

Heterogeneous distributed system (HDS) is made up of many computational resources with different disposal ability usually, and these computational resources are connected with each other through the express network that can satisfy various application requirements.In the past few years, heterogeneous distributed system has become popular computing platform, for computation-intensive and communications-intensive tasks provide various computation requirements.

Heterogeneous distributed system is used widely at numerous areas such as industry and commerce.Yet heterogeneous distributed system has also consumed huge energy when powerful calculating ability is provided.According to UN (Energy User News) report, now, cluster computing center energy requirement will be increased to 200-300W/ft2 in the coming years from 150W/ft2 to 200W/ft2.Environmental Protection Department of the United States Federal (EPA) has submitted a report about data center in August, 2009.This part report claims that the energy resource consumption of data center increased by one times from 2000 to 2006, expected 2011 and doubled.Huge energy consumption cost becomes the important bottleneck of the heterogeneous distributed system development of restriction, is badly in need of effectively being solved.

Represent that with directed acyclic graph DAG (Directed Acyclic Graph) the Parallel Task Scheduling problem is the research focus of academia always, the Parallel Task Scheduling problem has been proved to be np complete problem, has caused extensive studies at home and abroad.The target of scheduling is will be under the prerequisite that satisfies the priority of task restriction relation; Confirm a kind of assignment and execution sequence according to the suitable dispensing strategy; But the task reasonable distribution of executed in parallel is carried out to each processor in an orderly manner, to reach the purpose that reduces total execution time.

Based on different scheduling strategies, existing dispatching method mainly can be divided three classes: list scheduling method, the dispatching method that clusters, the dispatching method that duplicates based on task.Energy-conservation in the heterogeneous distributed system can be taked traditional dispatching method, and operational scheme is also closely similar, just when scheduling, taken all factors into consideration energy-conservation this target.Wherein, the selection of dispatching method has bigger influence for scheduling result, because inappropriate dispatching method may make real powerful computing ability can not get embodying, even can reduce system performance, increases system energy consumption.At present, existing energy-saving scheduling method great majority are based on above three kinds of dispatching methods, combine dynamic electric voltage adjusting (DVS), dynamic power management energy-conservation technology such as (DPM) to come the optimization system energy consumption again.

It is a breakthrough of microprocessor low power dissipation design in recent years that dynamic electric voltage is regulated (DVS) technology, and it allows dynamically to regulate microprocessor voltage and frequency, guarantees system's operate as normal simultaneously.At present, in the embedded hardware circuit design, no matter be CPU, still peripheral I/O interface all generally uses the CMOS logical circuit.In the CMOS logical circuit, power consumption is directly proportional with the quadratic power of frequency and voltage.Though reduce the increase that frequency/voltage can cause task execution time; But explore task free time according to the dependence between the parallel task; Utilize dynamic electric voltage to regulate (DVS) reasonable in technology ground and regulate processor voltage; Under the prerequisite that does not influence other tasks execution, can reduce the processor energy consumption effectively.In the voltage-regulation process, the voltage management chip can be carried out the adjustment of small voltage, and can be in the extremely short time adjustment of (tens microseconds in) completion voltage.

At present, mainly contain about the energy-saving scheduling method under the heterogeneous distributed system environments both at home and abroad:

The people such as D.P.Agrawal of U.S. University of Cincinnati have proposed a kind of dispatching algorithm of duplicating based on task (TDS) in heterogeneous system, this algorithm as much as possible the replication task forerunner to improve system performance.Then, proposed a kind of dispatching algorithm based on heterogeneous network (TANH) again, this algorithm has carried out improving with further sophisticated systems performance on the basis of TDS algorithm.

The dispatching method that duplicates based on task through duplicating certain or some tasks to nonidentical processor, is eliminated the communication overhead between the task with restriction relation, thereby is reduced whole task scheduling time span exactly.This method can reduce task scheduling length effectively, saves the communication energy consumption, but also can increase the calculating energy consumption of replication task simultaneously, therefore in energy-saving distribution, needs careful consideration.

Ancestor's of U.S. Auburn University is good to wait the people to propose a kind of energy-saving distribution algorithm that duplicates based on task, and this algorithm has proposed a threshold value and come control task to duplicate, and promptly optionally the best forerunner of replication task comes balance system performance and energy consumption two aspects.People such as the Gong Bin of Shandong University have proposed a kind of dispatching method and task of will clustering and have duplicated the power-economizing method that dispatching method effectively combines, and utilize dynamic power management (DPM) technology to save the idle energy consumption of processor.

The Lee of Sydney University has proposed a kind of heuristic dispatching method of regulating (DVS) technology based on dynamic electric voltage; This method considers in the task scheduling process aspect performance and the processor energy consumption two simultaneously, proposed a relative priority level (RS) as evaluation index set the tasks dispatch processor and the execution voltage on this processor.The Kang Yan of Yunnan University has proposed a kind of power-economizing method in heterogeneous distributed system; This method at first utilizes taboo (Tabu) strategy to obtain free time, utilizes dynamic electric voltage (DVS) technology of regulating to reduce the voltage of current task on processor to save the CPU energy consumption then.People such as the Kang of Univ Florida USA have proposed a kind of scheduling strategy based on dynamic electric voltage adjusting (DVS) technology, through effective allocating task reasonable adjustment of free time processor voltage, thus the minimizing processor energy consumption.

In sum, in the past most of certain part (system energy consumption comprises processor energy consumption and network energy consumption two parts) to performance or energy consumption of dispatching method has carried out improving or optimization.The weak point that exists has: the method that 1. has is only considered performance and is ignored energy consumption fully; 2. some energy-saving scheduling method that duplicates based on task utilizes threshold value that task is duplicated to control, but given threshold value be to be provided with arbitrarily, can not regulate according to parallel task and system environments self-adaptation, cause the scheduling result instability; Though the method that 3. has has not only been considered performance but also considered the processor energy consumption, has ignored the network service energy consumption.

Summary of the invention

The purpose of this invention is to provide in a kind of heterogeneous distributed system energy-saving scheduling method based on adaptive threshold; This method will be duplicated dispatching method and dynamic electric voltage (DVS) technology of regulating based on the task of adaptive threshold and combined; Under the prerequisite that satisfies the system performance requirement, can save energy consumption effectively.

Technical scheme of the present invention may further comprise the steps:

One, at first reads parallel task directed acyclic graph DAG file.

Two, obtain initiating task scheduling sequence.For parallel task collection V, total n=|V| task is from export task v _nBeginning, the priority of calculating each task is until beginning task v ₁Finish.Arrange according to task priority size ascending order then, obtain initiating task scheduling sequence.

Three, obtain optimal threshold.At first, first task of dispatching sequence from initiating task begins to travel through all tasks, calculates and duplicates current task v _iBest forerunner (FP (v _i)) the energy consumption moreenergy that increased (equals FP (v _i) the calculating energy consumption deduct it and v _iBetween the communication energy consumption), get wherein minimum value and maximal value respectively as minimum threshold min_threshold and max-thresholds max_threshold.Task scheduling length when the dispatching method that then, utilizes this patent and proposed asks threshold value to be respectively max_threshold and min_threshold-1 is minimum scheduling length and maximum scheduling length.At last; Specify one can satisfy the scheduling length that user performance requires; Begin to travel through all threshold values from min_threshold-1; And utilize dispatching method to ask task scheduling length under the current threshold value, less than specifying scheduling length just to finish traversal, current threshold value is optimal threshold up to the scheduling length of being tried to achieve.

Four, task is divided into groups.Begin from first task of initiating task scheduling sequence, carry out depth-first search up to the beginning task termination.In the task search procedure, if current task v _iBest forerunner (FP (v _i)) divide into groups, then it is distributed to the grouping at current task place, and be labeled as and distribute, otherwise to whether duplicating FP (v _i) carry out.If duplicate FP (v _i), then can increase replication task FP (v _i) the calculating energy consumption, reduce task scheduling length and FP (v _i) and v _iBetween the communication energy consumption.Suppose that the time that reduces is lesstime (the scheduling length that equals to reduce), the final energy consumption that increases is that moreenergy (equals FP (v _i) the calculating energy consumption deduct it and v _iBetween the communication energy consumption), if lesstime＞0, and moreenergy is smaller or equal to optimal threshold, then with replication task FP (v _i) add current task v to _iThe grouping at place, otherwise withdraw from current group and choose in the initiating task scheduling sequence first and do not divide group task to carry out next one grouping.

Five, duty mapping.From first beginning of dividing into groups, calculate and respectively be grouped in energy consumption on all unappropriated processors, current group is mapped on the minimum processor of energy consumption, and this processor of mark to be occupied, circulation like this is gone down and all is assigned with away up to all groupings.

Six, processor voltage is regulated.Utilize the free time that dependence produced between the task, the execution time under each voltage on the dispatch processor of setting the tasks, make that the calculating energy consumption that produces is minimum, then dynamic adjustment processor voltage in the task implementation.

Characteristics of the present invention are to the energy-optimised problem in the heterogeneous distributed system, will combine optimization system energy consumption when satisfying the system performance requirement based on task replication strategy and dynamic electric voltage (DVS) technology of regulating of adaptive threshold.At first, can dynamically obtain an optimal threshold according to parallel task and system environments (mainly comprise system provided processor and network parameter) based on the task replication strategy of adaptive threshold among the present invention, improve the dirigibility of system; Utilize this threshold value that control task is duplicated then, promptly optionally the best forerunner of replication task comes balance system performance and energy consumption to divide into groups to obtain near-optimization, has improved the stability of system; At last; With each packet scheduling on unappropriated and processor that energy consumption is minimum; This processor utilizes dynamic electric voltage (DVS) technology of regulating dynamically to adjust processor voltage with further saving processor energy consumption according to the free time that dependence produced between the task.

Description of drawings

Accompanying drawing is a FB(flow block) of the present invention.

Embodiment

Below in conjunction with accompanying drawing the present invention is made further detailed description.

Parallel task: the task that the user submits to representes with directed acyclic graph DAG, be defined as G (V, E).V=(v wherein ₁, v ₂, L, v _n) expression forgives the task-set of n task.For each task among the V, task v _iAt processor p _jOn execution time be t _Ij, wherein maximum execution time is designated as

t _i ^*Expression task v _iExecution time on the processor that is scheduled.Set forth for ease, use { V here respectively ₁, V ₂, K, V _hAnd { f ₁, f ₂, K, f _hRepresent that each task belongs to the voltage and the frequency set of processor.Especially, as task v _iWhen having free time, can utilize dynamic electric voltage to regulate the execution voltage on the processor of (DVS) technological adjusting task place, task is at voltage V _kUnder execution time be designated as τ _IkIn addition, for task v _i, task computation cycle cc _iBe certain value, irrelevant with processor voltage.E is a message set, e _Ij=(v _i, v _j) ∈ E representes task v _iTo task v _jMessage transmitted, c _IjExpression pass-along message e _IjThe call duration time of the expense of sending out.In addition, we use succ (v respectively _i) and pred (v _i) expression task v _iThe set of follow-up set of tasks and predecessor task.

System resource: a heterogeneous distributed system is made up of many processor nodes that can have the different disposal ability, is designated as P={p ₁, p ₂, K, p _m, these nodes are connected with each other through express network Link.All processors in the system all support dynamic electric voltage to regulate (DVS) technology, and each processor all has h the magnitude of voltage of arranging by descending order, are designated as { V _I1, V _I2, K, V _Ih, frequency value corresponding is designated as { f _I1, f _I2, K, f _Ih, 1≤i≤n wherein.We can use the two-dimensional matrix X of a n * m to represent the scheduling process of n duty mapping to m heterogeneous processor.If task v _iBe dispatched to processor p _j, x then _Ij=1, otherwise x _Ij=0, x wherein _Ij∈ X.

Embodiment is carried out according to following steps.

One, reads parallel task directed acyclic graph DAG file.

Two, obtain initiating task scheduling sequence.For parallel task collection V, total n=|V| task is from export task v _nBeginning, the priority of calculating each task is up to beginning task v ₁Finish.Arrange according to task priority size ascending order then, obtain initiating task scheduling sequence.Wherein, task v _iPriority bottom (v _i) define shown in formula (1).

$\mathrm{bottom}\left(v_i\right)=\left\{\begin{array}{c}\stackrel{\‾}{t_i}\mathrm{ifsucc}\left(v_i\right)=\mathrm{\Φ},\\ \underset{v_k\∈\mathrm{succ}\left(v_i\right)}{\max }\left(\mathrm{bottom}\left(v_k\right)\right)+\stackrel{\‾}{t_i},\mathrm{otherwise}.\end{array}\right.---\left(1\right)$

Three, obtain optimal threshold.At first, calculated threshold span.Begin to travel through all tasks from first task of initiating task scheduling sequence, calculate and duplicate current task v _iBest forerunner (FP (v _i)) the energy consumption moreenergy that increased (equals FP (v _i) the calculating energy consumption deduct it and v _iBetween the communication energy consumption), i.e. moreenergy=en _k-el _Ki(represent FP (v with k here _i)), minimum value and the maximal value of getting its moreenergy are respectively as minimum threshold min_threshold and max-thresholds max_threshold.Wherein best forerunner FP (v _i) define shown in formula (2):

FP(v _i)＝v _k，s.t.ECT(v _k)+c _ki≥ECT(v _l)+c _li?where?v _k∈pred(v _i)，v _l∈pred(v _i)&&l≠k.(2)

Task v _iEarliest start time on each processor (EST) is recursive calculation from top to bottom, and the earliest start time of the task that wherein enters the mouth is 0.Computing formula is:

$\mathrm{EST}(v_i/w)=\left\{\begin{array}{c}0,\mathrm{ifpred}\left(v_i\right)=\mathrm{\Φ}\\ \underset{e_{\mathrm{li}}\∈E}{\min }\left(\underset{\mathrm{FP}\left(v_i\right)=w}{\underset{e_{\mathrm{kj}}\∈E,v_k\≠v_l}{\max }}(\mathrm{ECT}\left(v_k\right),\mathrm{ECT}\left(v_l\right)+c_{\mathrm{li}})\left|\right|\underset{\mathrm{FP}\left(v_i\right)\≠w}{\underset{e_{\mathrm{ki}}\∈E,v_k\≠v_l}{\max }}(\mathrm{EST}\left(v_k\right)+t_{\mathrm{kw}},\mathrm{ECT}\left(v_l\right)+c_{\mathrm{li}})\right),\mathrm{otherwise}\end{array}\right.---\left(3\right)$

Here, task v _iActual earliest finish time can use task v _iAt its optimum processor (fproc (v _i)) on earliest finish time represent:

EST(v _i)＝EST(v _i/fproc(v _i)) (4)

Task v _iEarliest finish time (ECT) equal task v _iActual earliest start time add task v _iExecution time on its optimum processor, shown in formula (5):

ECT(j)＝EST(j)+t _jfproc(j). (5)

Task v _iOptimum processor can calculate according to formula (6):

$\mathrm{fproc}\left(v_i\right)=p\&Element;\mathrm{Ps}.t.\&ForAll;q\&Element;P,(\mathrm{EST}(v_i/p)+t_{\mathrm{ip}}<\mathrm{EST}(v_i/q)+t_{\mathrm{iq}})---\left(6\right)$

For task v _i, it is at processor p _jOn the calculating power consumption with formula (7) expression, wherein PN _JkBe processor p _jAt voltage V _kUnder power consumption.

${\mathrm{en}}_i=\underset{k=1}{\overset{h}{\mathrm{\&Sigma;}}}{\mathrm{PN}}_{\mathrm{jk}}{\mathrm{\&tau;}}_{\mathrm{ik}}---\left(7\right)$

Like this, the activity energy consumption of all tasks can be used formula (8) expression.Wherein, PN _{I_highest}, f _{I_highest}, V _{I_highest}Be respectively task v _iMaximum power dissipation, maximum frequency and the maximum voltage of place processor.In addition, all processor free time can be consumed formula (9) expression, and makespan is the scheduling length of the parallel task after task scheduling finishes.

${\mathrm{EN}}_{\mathrm{active}}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{h}{\mathrm{\&Sigma;}}}{\mathrm{PN}}_{i\_\mathrm{highest}}\&CenterDot;\frac{f_j{V}_j^2}{f_{i\_\mathrm{highest}}{V}_{i\_\mathrm{highest}}^2}{\mathrm{\&tau;}}_{\mathrm{ij}}---\left(8\right)$

${\mathrm{EN}}_{\mathrm{idle}}=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}({\mathrm{PN}}_{i\_\mathrm{idle}}\&CenterDot;(\mathrm{makespan}-\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{h}{\mathrm{\&Sigma;}}}(x_{\mathrm{ij}}\&CenterDot;{\mathrm{\&tau;}}_{\mathrm{jk}})))---\left(9\right)$

Message transfer e _Ij=(v _i, v _j) energy that consumed of ∈ E can use formula (10) to represent.Like this, the network service energy consumption EL in the total system can use formula (11) expression.Wherein, PL is the network service power consumption.

el _ij＝PL×c _ij (10)

$\mathrm{EL}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\underset{j=1,j\&NotEqual;i}{\overset{n}{\mathrm{\&Sigma;}}}\underset{p=1}{\overset{m}{\mathrm{\&Sigma;}}}\underset{q=1,q\&NotEqual;p}{\overset{m}{\mathrm{\&Sigma;}}}(x_{\mathrm{ip}}\&CenterDot;x_{\mathrm{jq}}\&CenterDot;\mathrm{PL}\&CenterDot;c_{\mathrm{ij}})---\left(11\right)$

Secondly, calculate scheduling length span.Task scheduling length when the dispatching method that utilizes this patent to propose is asked threshold value for max_threshold and min_threshold-1 respectively is minimum scheduling length and maximum scheduling length.The scheduling length calculation comprises following three steps: the first step, and under current threshold value, utilize the grouping strategy in the step 4 that task is divided into groups; Second step, utilize scheduling strategy in the step 5 with duty mapping to the corresponding processing device; In the 3rd step, obtain the task scheduling length under the current threshold value according to the task scheduling result.Details can be referring to step 4 and step 5.

At last; Specify one can satisfy the scheduling length that user performance requires; All threshold values of traversal from min_threshold-1 to max_threshold; The dispatching method that utilizes this patent to propose is found the solution the task scheduling length under the current threshold value, is less than or equal to up to the scheduling length of being tried to achieve and specifies scheduling length just to finish traversal, and the current threshold value of this moment is optimal threshold.

Four, task is divided into groups.Begin from first task of initiating task scheduling sequence, carry out depth-first search task is divided into groups.In the task search procedure, if current task v _iBest forerunner (FP (v _i)) do not divide into groups, then it is assigned to v _iThe grouping at place, and be labeled as and distribute, otherwise to whether duplicating FP (v _i) weigh.If duplicate FP (v _i), then can increase replication task FP (v _i) the calculating energy consumption, reduce task scheduling time span and FP (v _i) and v _iBetween the communication energy consumption, the time of minimizing is designated as lesstime, the energy consumption of increase is designated as moreenergy.If lesstime＞0, and moreenergy is less than or equal to optimal threshold, then with replication task FP (v _i) add current task v to _iThe grouping at place, otherwise withdraw from current group and choose in the initiating task scheduling sequence first and do not divide group task to carry out next one grouping.Wherein, reduce time lesstime=LACT (v _i)-LAST (v _j)+c _Ij, increase energy consumption moreenergy=en _k-el _Ki

The deadline of permission the latest (LACT) of export task equals its earliest finish time, and the deadline of permission the latest of other tasks is recursive calculation from top to bottom, shown in formula (12):

$\mathrm{LACT}\left(v_i\right)=\left\{\begin{array}{c}\mathrm{ECT}\left(v_i\right),\mathrm{ifsucc}\left(v_i\right)=\mathrm{\&Phi;}\\ \min (\underset{e_{\mathrm{ik}}\&Element;E,v_i\&NotEqual;\mathrm{FP}\left(v_k\right)}{\min }(\mathrm{LAST}\left(v_k\right)-c_{\mathrm{ik}}),\underset{e_{\mathrm{ik}}\&Element;E,v_i=\mathrm{FP}\left(v_k\right)}{\min }\left(\mathrm{LAST}\left(v_k\right)\right)),\mathrm{otherwise}.\end{array}\right.---\left(12\right)$

Task v _iThe start time of permission the latest (LAST) equal it and allow the deadline to deduct its execution time the latest, LAST (v _i) define like formula (13):

$\mathrm{LAST}\left(v_i\right)=\mathrm{LACT}\left(v_i\right)-t_{\mathrm{ifproc}\left(v_i\right)}.---\left(13\right)$

Five, duty mapping.From first grouping beginning; Calculate and respectively be grouped in the energy consumption on all unappropriated processors; Current group is mapped on the minimum processor of energy consumption, and this processor of mark is occupied, so circulation is gone down till all task groupings all are assigned with away.

Six, processor voltage is regulated.Utilize the dependence exploration task free time between the task, the execution time that sets the tasks under each voltage on the dispatch processor makes the task computation energy consumption minimum, then dynamic adjustments processor voltage in the task implementation.

For task v _i, its computation period and execution time, the relation of carrying out between the frequency are: cc _i=t _i ^*F _iIf task v _iAlso there is free time after under processor highest frequency/voltage, being finished, then can saves the calculating energy consumption through the voltage/frequency that reduces on its place processor.At this moment, computation period can be expressed as:

${\mathrm{cc}}_i=\underset{j=1}{\overset{h}{\mathrm{\&Sigma;}}}{\mathrm{\&tau;}}_{\mathrm{ij}}\&CenterDot;f_j---\left(14\right)$

Here, { τ _Ij, i=1, L, n, j=1, L, h} ∈ Integer (15)

According to the dependence between the task, we can obtain duty mapping to processor actual beginning execution time (AST) and maximum allowable execution time (MAET) afterwards, and account form is shown in formula (16) and formula (17).Wherein, group (v _i=v _j) expression task v _iAnd v _jIn same grouping.

$\mathrm{AST}\left(v_i\right)=\left\{\begin{array}{c}0,\mathrm{ifpred}\left(v_i\right)=\mathrm{\&Phi;}\\ \max (\underset{v_j\&Element;\mathrm{pre}\left(v_i\right),\mathrm{group}(v_i\&NotEqual;v_j)}{\max }(\mathrm{AST}\left(v_j\right)+t_j^{*}+c_{\mathrm{ji}}),\underset{v_j\&Element;\mathrm{pre}\left(v_i\right),\mathrm{group}(v_i=v_j)}{\max }(\mathrm{AST}\left(v_j\right)+t_j^{*})),\mathrm{otherwise}\end{array}\right.---\left(16\right)$

$\mathrm{MAET}\left(v_i\right)=\left\{\begin{array}{c}{t}_i^{*},\mathrm{ifsucc}\left(v_i\right)=\mathrm{\&Phi;}\\ \min (\underset{v_j\&Element;\mathrm{succ}\left(v_i\right),\mathrm{group}(v_i\&NotEqual;v_j)}{\min }(\mathrm{AST}\left(v_j\right)-\mathrm{AST}\left(v_i\right)-c_{\mathrm{ij}}),\underset{v_j\&Element;\mathrm{succ}\left(v_i\right),\mathrm{group}(v_i=v_j)}{\min }(\mathrm{AST}\left(v_j\right)-\mathrm{AST}\left(v_i\right))),\mathrm{otherwise}.\end{array}\right.---\left(17\right)$

Therefore, task v _iExecution time satisfy below relation:

$t_i^{*}\&le;\underset{j=1}{\overset{h}{\mathrm{\&Sigma;}}}{\mathrm{\&tau;}}_{\mathrm{ij}}\&le;{\mathrm{MAET}}_i---\left(18\right)$

The purpose that adopts dynamic electric voltage to regulate (DVS) technology among the present invention is to minimize the task computation energy consumption through the execution time of exploration task under each voltage of processor.Here, we had both considered processor activity energy consumption, had considered the idle energy consumption of processor again.Therefore, the target equation of each task can be expressed as:

$\mathrm{Min}{\mathrm{en}}_i={\mathrm{en}}_{\mathrm{active}}^i+{\mathrm{en}}_{\mathrm{idle}}^i=\underset{j=1}{\overset{h}{\mathrm{\&Sigma;}}}{\mathrm{PN}}_{i\_\mathrm{highest}}\&CenterDot;\frac{f_j{V}_j^2}{f_{\mathrm{highest}}{V}_{\mathrm{highest}}^2}{\mathrm{\&tau;}}_{\mathrm{ij}}+{\mathrm{PN}}_{i\_\mathrm{idle}}\&CenterDot;({\mathrm{MAET}}_i-\underset{j=1}{\overset{h}{\mathrm{\&Sigma;}}}{\mathrm{\&tau;}}_{\mathrm{ij}})---\left(19\right)$

Above optimization problem is an integral linear programming problem; It is an equation of condition with formula (14), (15), (18); With formula (19) is the target equation, can utilize integral linear programming to find the solution instrument LP_solve and try to achieve the integer optimum solution of task in each following execution time of voltage of processor.

Claims (1)

1. In the heterogeneous distributed system based on the energy-saving scheduling method of adaptive threshold, it is characterized in that: may further comprise the steps:

   One, at first reads parallel task directed acyclic graph DAG file;

   Two, obtain initiating task scheduling sequence; For parallel task collection V, total n=|V| task is from export task v _nBeginning, the priority of calculating each task is until beginning task v ₁Finish; Arrange according to task priority size ascending order then, obtain initiating task scheduling sequence;

   Three, obtain optimal threshold; At first, first task of dispatching sequence from initiating task begins to travel through all tasks, calculates and duplicates current task v _iBest forerunner (FP (v _i)) the energy consumption moreenergy that increased (equals FP (v _i) the calculating energy consumption deduct it and v _iBetween the communication energy consumption), get wherein minimum value and maximal value respectively as minimum threshold min_threshold and max-thresholds max_threshold; Task scheduling length when the dispatching method that then, utilizes this patent and proposed asks threshold value to be respectively max_threshold and min_threshold-1 is minimum scheduling length and maximum scheduling length; At last; Specify one can satisfy the scheduling length that user performance requires; Begin to travel through all threshold values from min_threshold-1; And utilize dispatching method to ask task scheduling length under the current threshold value, less than specifying scheduling length just to finish traversal, current threshold value is optimal threshold up to the scheduling length of being tried to achieve;

   Four, task is divided into groups; Begin from first task of initiating task scheduling sequence, carry out depth-first search up to the beginning task termination; In the task search procedure, if current task v _iBest forerunner (FP (v _i)) divide into groups, then it is distributed to the grouping at current task place, and be labeled as and distribute, otherwise to whether duplicating FP (v _i) carry out; If duplicate FP (v _i), then can increase replication task FP (v _i) the calculating energy consumption, reduce task scheduling length and FP (v _i) and v _iBetween the communication energy consumption; Suppose that the time that reduces is lesstime (the scheduling length that equals to reduce), the final energy consumption that increases is that moreenergy (equals FP (v _i) the calculating energy consumption deduct it and v _iBetween the communication energy consumption), if lesstime＞0, and moreenergy is smaller or equal to optimal threshold, then with replication task FP (v _i) add current task v to _iThe grouping at place, otherwise withdraw from current group and choose in the initiating task scheduling sequence first and do not divide group task to carry out next one grouping;

   Five, duty mapping; From first beginning of dividing into groups, calculate and respectively be grouped in energy consumption on all unappropriated processors, current group is mapped on the minimum processor of energy consumption, and this processor of mark to be occupied, circulation like this is gone down and all is assigned with away up to all groupings;

   Six, processor voltage is regulated; Utilize the free time that dependence produced between the task, the execution time under each voltage on the dispatch processor of setting the tasks, make that the calculating energy consumption that produces is minimum, then dynamic adjustment processor voltage in the task implementation.

Images