CN103942102A - Real-time task scheduling method based on dual priority - Google Patents

Abstract

The invention discloses a real-time task scheduling method based on dual priority. The method comprises the steps that firstly, the lengths ransksp (ti) of paths from tasks to an entrance task are calculated, the lengths are arranged according to the ascending sequence, and a scheduling queue is obtained; then, the length rankup (ti) of the path from a task ti to an exit task is calculated; the tasks with the identical ranksp (ti) are arranged according to the rankup (ti) descending sequence and the rankup (ti) obtained through calculation, and the scheduling queue is updated; a task priority scheduling queue is constructed until no tasks with the identical ranksp (ti) exist; finally, task resources are selected according to the scheduling queue. The thought of a task interval interposition and list scheduling algorithm is combined, the real-time task scheduling method DPSA based on the dual priority is provided, and the scheduling sequence is arranged with the calculating method of the lengths of the paths from the tasks to the entrance task and improved HEFT weights as the priority so that task scheduling can be more reasonable.

Description

Based on the real-time task scheduling method of double priority level

Technical field

The present invention relates to a kind of method for scheduling task, be specifically related to a kind of real-time task scheduling method based on double priority level.

Background technology

Task is to be reasonably allocated in corresponding resource according to certain Rule and constraint condition, and the order of executing the task is to be controlled by task scheduling algorithm.The target of task scheduling is communication overhead and the stand-by period between minimizing task, makes the execution time of task and deadline the shortest, to improve the performance of system.

Task scheduling is mainly divided into two kinds of static scheduling and dynamic dispatchings.Scheduling problem generally uses directed acyclic graph (Directed acyclic graph, DAG) to represent.The method of Static task scheduling mainly contains: based on the scheduling of didactic scheduling and intelligent search, because Mission Scheduling is a np complete problem, conventionally adopt heuristic dispatching method to solve this class problem.Heuristic dispatching algorithm mainly comprises: the dispatching algorithm based on sub-clustering, dispatching algorithm and the List scheduling algorithm based on replication strategy.List scheduling algorithm is widely studied with its simple thought and higher dispatching efficiency, up to the present a lot of algorithms have been there are, more existing classical List scheduling algorithms have: HEFT(Heterogeneous earliest finish time, isomery earliest finish time), CPOP(Critical-path-on-a-processor, critical path on processor), LDCP(Longest dynamic critical path, the longest Dynamic Critical Path) etc.A lot of algorithms are all to improve research based on these algorithms, and that wherein improves and study based on HEFT algorithm is maximum.

HEFT algorithm is the algorithm based on list scheduling, and this algorithm mainly comprises two stages: task priority phase sorting and resource selection stage.In the task prioritization stage, calculate the longest path of each task to export task, and by calculating income value descending sort, but owing to may there is the task that multiple values are equal, the task of now choosing is at random dispatched; Be task allocated phase in the resource selection stage, on the earliest finish time while assigning the task to each resource based on interval insertion technique computes, choose the resource of minimum on earliest finish time as the scheduling resource of task.

Document 1[Daoud M I, Kharma N.A high performance algorithm for static task scheduling in heterogeneous distributed computing systems[J]。Journal of Parallel and distributed computing, 2008,68(4): 399-409] one is proposed based on the longest Dynamic Critical Path algorithm LDCP(Longest Dynamic Critical Path Algorithm).Document 2[Jiang Yun connection, Sun Guangzhong, Xu Yinlong.The efficient task scheduling algorithm of one [J] in Heterogeneous Parallel Systems.Computer engineering, 2007,11:39-41] proposed a kind ofly to insert and the efficient heuritic approach HDEFT(Heterogeneous Duplication based Earliest Finish Time of Task Duplication based on interval), it by calculation task to the weights of its all subsequent tasks with decide the priority of task.Document 3[Zhou L, Shixin S.Scheduling algorithm based on critical tasks in heterogeneous environments[J]。Systems Engineering and Electronics, Journal of, 2008,19(2): 398-404] crucial task scheduling algorithm HCT(Heterogeneous critical task under a kind of isomerous environment is proposed), this algorithm copies to mission critical the resource having on free time sheet, ahead of time the task start time.Document 4[wishes a treasure, Xiao Dan.Task scheduling algorithm [J] based on path priority under cloud computing environment.Computer engineering and design, 2013,10:3511-3515] task scheduling algorithm based on path priority under the cloud computing environment that proposes, its drawing-in system isomery parameter is calculated the weights of task and limit in DAG figure, can reflect more exactly the priority level of task.The little pleasure of document 5[king, Huang Hongbin, Deng Su.The information physics emerging system static task List scheduling algorithm [J] of processing sequence constraint.Robotization journal.2012,38(11): 1870-1879] the IHEFT(Improvement Heterogeneous Earliest Finish Time that proposes) algorithm is the up weighing computation method that improves HEFT algorithm, obtains more rational task scheduling sequencing queue.

Most List scheduling algorithm only adopts the priority of single property calculation task, and in the time that priority equates, random selection task is dispatched, and does not consider that the method for choosing at random task is unfavorable for the priority scheduling of vital task.

Summary of the invention

Technical matters to be solved by this invention is to be the priority with single property calculation task for current most of List scheduling algorithm, in the time that priority equates, random selection task is dispatched, do not consider that the method for choosing at random task is unfavorable for the priority scheduling of vital task, a kind of real-time task scheduling method based on double priority level proposing.

For addressing the above problem, the present invention adopts following scheme to realize:

Based on a real-time task scheduling method for double priority level, comprise two stages of resource selection that build task priority dispatching queue stage and task.

(1) build the task priority dispatching queue stage:

(1.1) according to formula calculate each task t _ito the path rank of entrance task _sp(t _i), and arrange by its ascending order, obtain scheduling queue;

${\mathrm{rank}}_{\mathrm{sp}}\left(t_i\right)=\underset{r_k\∈R}{\min }\underset{t_j\∈\mathrm{pred}\left(t_i\right)}{\max }\{{\mathrm{rank}}_{\mathrm{sp}}\left(t_j\right)+w_{j,k}+{c^{\′}}_{\mathrm{ji}}\}$

In formula, rank _sp(t _i) expression task t _ito the path of entrance task, rank _sp(t _j) expression task t _jto the path of entrance task, t _jt _ipredecessor task, rank _sp(t _entry)=0; w _j,kexpression task t _jat resource r _kon processing expenditure, R={r ₁, r ₂, r _mthe set of expression heterogeneous resource, pred(t _i) expression task t _ipredecessor task set; C ' _jiexpression task t _jto task t _ipractical communication expense, r _k=succ_R (t _i) expression task t _jthe resource r choosing _kwith task t _ipre-allocation resource identical;

(1.2) according to formula calculate each task t _ito the path rank of export task _up(t _i);

${\mathrm{rank}}_{\mathrm{up}}\left(t_i\right)=\underset{r_k\∈R}{\min }\underset{t_j\∈\mathrm{succ}\left(t_i\right)}{\max }\{{\mathrm{rank}}_{\mathrm{up}}\left(t_j\right)+w_{i,k}+c_{\mathrm{ij}}^{\′}\}$

In formula, rank _up(t _i) expression task t _ito the path of export task, rank _up(t _j) expression task t _jto the path of export task, t _jt _isubsequent tasks, w _i,kexpression task t _iat resource r _kon processing expenditure, R={r ₁, r ₂, r _mthe set of expression heterogeneous resource, succ(t _i) expression task t _isubsequent tasks set; C ' _ijexpression task t _ito task t _jpractical communication expense, r _k=pred_R (t _i) expression task t _jthe resource r choosing _kwith task t _ipre-service resource identical; J=1,2,3 ..., n; K=1,2 ... m;

(1.3) if there is identical rank _sp(t _i) task, to these rank _sp(t _i) the rank that calculates according to step (1.2) of task _up(t _i), by rank _up(t _i) descending sort, upgrade scheduling queue;

(1.4) until there is not identical rank _sp(t _i) task, task priority dispatching queue has built;

Above-mentioned j=1,2 ... n; I=1,2 ... n, n represents the sum of task; K=1,2 ... m, m represents the sum of heterogeneous resource;

(2) resource selection of task: according to the priority scheduling queue of the task of above-mentioned structure, the processing resource to task is selected.

In step (1.1), if there are multiple entrance tasks in set of tasks, need to build a virtual entrance task, this virtual entrance task is 0 to the communication overhead between each task; In step (1.2), if there are multiple export tasks in set of tasks, need to build a virtual export task, this virtual export task is 0 to the communication overhead between each export task.

For each task, only have after its predecessor task completes and could carry out.

The communication overhead of carrying out in same resource of task is 0.

Described step (2) is specially, according to formula with can calculate start time and the deadline of task, to the resource selection of task, the deadline that selection can make task resource the earliest;

The start time of task

$\mathrm{est}(t_i,r_k)=\underset{t_j\∈\mathrm{pred}\left(t_i\right)}{\max }\{\mathrm{eft}(t_j,r_m)+c_{\mathrm{ji}}^{\′}\}$

Wherein $c_{\mathrm{ji}}^{\′}=\left\{\begin{array}{cc}{c}_{\mathrm{ji}},& m\≠k\\ 0,& m=k\end{array}\right..$

The deadline of task

eft(t _i,r _k)=est(t _i,r _k)+w _i,k

Wherein est (t _entry, r)=0.

Compared with prior art, the present invention is in conjunction with the thought of insertion and List scheduling algorithm between mission area, a kind of real-time task scheduling method DPSA(Double Priority Scheduling Algorithm based on double priority level is proposed), it adopts task to arrange the dispatching sequence of task as its priority to the path of entrance task and the computing method of improved HEFT weights, so that the scheduling of task is more reasonable.

Brief description of the drawings

Fig. 1 is the DAG structural representation of the real-time task scheduling method based on double priority level.

Fig. 2 is real-time task scheduling method based on double priority level and the scheduling length comparison of HEFT method.Embodiment

Based on a real-time task scheduling method for double priority level, comprise the steps:

1, task model

With DAG model G(T, the E of Weighted Coefficients) dependence between expression task.As shown in Figure 1.Wherein T represents the set of task, T (t ₁, t ₂, t _n) representing task-set, n represents the sum of task.E is the directed edge of Weighted Coefficients, communication set and dependence between expression task, E{e _ij| e _ij=<t _i, t _j>, e _ij∈ T × T}, e _ij∈ E represents task t _iwith t _jbetween dependence, and t _jmust be at t _iafter complete, could start to carry out, at this moment claim t _it _ja predecessor task, t _jt _isubsequent tasks, wherein entrance task does not have predecessor task, export task does not have subsequent tasks.

Owing to being in heterogeneous resource, R={r ₁, r ₂, r _mrepresent heterogeneous resource set, suppose that the communication speed between any two resources is identical here.W _i,kexpression task t _iat resource r _kon processing expenditure, k=1,2,3 ..., m, W=[w _i,k] _{n × m}the processing expenditure matrix of task in resource.

The present embodiment is made following provisions to DAG model:

1) if there are multiple entrance tasks, build a virtual entrance task, it is 0 to the communication overhead between each entrance task; If there are multiple export tasks, build a virtual export task, it is 0 to the communication overhead between each export task.

2), for each task, only have after its predecessor task completes and could carry out.

3) communication overhead of carrying out in same resource of task is 0.

The processing expenditure of table 1 task on different resource

The weights calculating of priority and the related definition of DAG:

1) task t _ito the path of entrance task

${\mathrm{rank}}_{\mathrm{sp}}\left(t_i\right)=\underset{r_k\&Element;R}{\min }\underset{t_j\&Element;\mathrm{pred}\left(t_i\right)}{\max }\{{\mathrm{rank}}_{\mathrm{sp}}\left(t_j\right)+w_{j,k}+{c^{\&prime;}}_{\mathrm{ji}}\}$ ①

Wherein rank _sp(t _entry)=0,

2) computing formula of the up weight of the improved HEFT algorithm IHEFT of the present embodiment reference,

The up weight calculation formula of HEFT algorithm:

${\mathrm{rank}}_{\mathrm{up}}\left(t_i\right)={\stackrel{\&OverBar;}{w}}_i+\underset{t_j\&Element;\mathrm{succ}\left(t_i\right)}{\max }\{{\mathrm{rank}}_{\mathrm{up}}\left(t_j\right)+c_{\mathrm{ij}}\}$ ②

Wherein represent average treatment expense.

The up weight calculation formula of IHEFT algorithm:

${\mathrm{rank}}_{\mathrm{up}}\left(t_i\right)=\underset{r_k\&Element;R}{\min }\underset{t_j\&Element;\mathrm{succ}\left(t_i\right)}{\max }\{{\mathrm{rank}}_{\mathrm{up}}\left(t_j\right)+w_{i,k}+c_{\mathrm{ij}}^{\&prime;}\}$ ③

Wherein ${\mathrm{rank}}_{\mathrm{up}}\left(t_{\mathrm{exit}}\right)=\underset{r_k\&Element;R}{\mathrm{mi}n}{w}_{\mathrm{exit},k},$

The present embodiment is considered rank _sp(t _i) possible identical situation, use rank _up(t _i) task is carried out to prioritization again.

3) calculating rank _sp(t _i) when weight, the resource of weight maximum is task t _ipre-allocation resource, with succ_R (t _i) represent, and there is rank _sp(t _i, succ_R (t _i))>=rank _sp(t _i, r).Calculating rank _up(t _i) when weight, the resource of weight minimum is the pre-allocation resource of task, with pred_R (t _i) represent. there is rank _up(t _i, pred_R (t _i))≤rank _up(t _i, r) set up.

4) start time of task

$\mathrm{est}(t_i,r_k)=\underset{t_j\&Element;\mathrm{pred}\left(t_i\right)}{\max }\{\mathrm{eft}(t_j,r_m)+c_{\mathrm{ji}}^{\&prime;}\}$ ⑤

Wherein $c_{\mathrm{ji}}^{\&prime;}=\left\{\begin{array}{cc}{c}_{\mathrm{ji}},& m\&NotEqual;k\\ 0,& m=k\end{array}\right..$

5) deadline of task

eft(t _i,r _k)=est(t _i,r _k)+w _i,k ⑥

Wherein est (t _entry, r)=0.

6) free time in resource

slot(t _i,r _k)=est(t _i,r _k)-ava(r _k) ⑦

Wherein ava (r _k) be resource r _kon available starting point.

7) scheduling length of task

makespan=max{eft(t _exit)} ⑧

2, build task priority dispatching queue

In the priority scheduling queue of the task of structure, first calculate rank _sp(t _i) and rank _up(t _i), according to rank _sp(t _i) carry out ascending sort, rank _sp(t _i) when identical, these tasks are pressed to rank _up(t _i) carry out descending sort, the final scheduling queue that must go out on missions according to this ranking results of twice.Algorithm steps is as follows:

(1) 1. calculate the rank of each task according to equation _sp(t _i), and arrange by its ascending order, obtain scheduling queue;

(2) 3. calculate the rank of each task according to equation _up(t _i);

(3) if there is identical rank _sp(t _i) task, to identical rank _sp(t _i) task calculate rank according to step (2) _up(t _i), by rank _up(t _i) descending sort, upgrade scheduling queue;

(4) until there is not identical rank _sp(t _i) task, task priority dispatching queue has built.

3, the resource selection of task

The priority scheduling queue of building according to a upper joint of task, we select the processing resource to task.This step is same as the prior art, according to formula 5. and 6. can calculate start time and the deadline of task, the resource of task is selected the deadline that selection can make task resource the earliest.Generally, the start time of a task is all after last task this resource completes, just to start to carry out, may exist like this two tasks processing in this resource to have enough free time to complete this task, we are just inserted into this task upper execution of free time in this resource, the task deadline ahead of time, thus scheduling length reduced.

Dispatching sequence, task deadline and the task of table 2 this method and HEFT algorithm are processed resource accordingly

Performance evaluation and result verification: the time complexity based on List scheduling algorithm is all lower, the time complexity of the present embodiment is divided into two parts: the priority scheduling queue of structure task and the resource selection of task, their time complexity is respectively O (n ²) and O (m × n ²), wherein n is task number, the number that m is resource, and the time complexity of algorithm is O (m × n herein ²), identical with the time complexity of HEFT algorithm, than the time complexity O (m × n of LDCP algorithm ³) little.The present embodiment is by change task number, and the maximum out-degree of the number of resource and task generates different DAG figure, when interstitial content is 5,15,20,30,50 o'clock, number of resources was 2,3,4,6,8, maximum out-degree is 3,5,8,10,14,16, obtain following result: as can be seen from Figure 2, the algorithm that the present invention provides is compared classic algorithm HEFT less scheduling length.

Claims (4)

1. the real-time task scheduling method based on double priority level, comprises two stages of resource selection that build task priority dispatching queue stage and task, it is characterized in that, the described structure task priority dispatching queue stage is specific as follows:

(1) build the task priority dispatching queue stage:

(1.1) according to formula calculation task is to the path rank of entrance task _sp(t _i), and arrange by its ascending order, obtain scheduling queue;

${\mathrm{rank}}_{\mathrm{sp}}\left(t_i\right)=\underset{r_k\&Element;R}{\min }\underset{t_j\&Element;\mathrm{pred}\left(t_i\right)}{\max }\{{\mathrm{rank}}_{\mathrm{sp}}\left(t_j\right)+w_{j,k}+{c^{\&prime;}}_{\mathrm{ji}}\}$

In formula, rank _sp(t _i) expression task t _ito the path of entrance task, rank _sp(t _j) expression task t _jto the path of entrance task, t _jt _ipredecessor task, rank _sp(t _entry)=0; w _j,kexpression task t _jat resource r _kon processing expenditure, R={r ₁, r ₂, r _mthe set of expression heterogeneous resource, pred(t _i) expression task t _ipredecessor task set; C ' _jiexpression task t _jto task t _ipractical communication expense, r _k=succ_R (t _i) expression task t _jthe resource r choosing _kwith task t _ipre-allocation resource identical;

(1.2) according to formula calculate each task t _ito the path rank of export task _up(t _i);

${\mathrm{rank}}_{\mathrm{up}}\left(t_i\right)=\underset{r_k\&Element;R}{\min }\underset{t_j\&Element;\mathrm{succ}\left(t_i\right)}{\max }\{{\mathrm{rank}}_{\mathrm{up}}\left(t_j\right)+w_{i,k}+c_{\mathrm{ij}}^{\&prime;}\}$

In formula, rank _up(t _i) expression task t _ito the path of export task, rank _up(t _j) represent task t in resource _jto the path of export task, t _jt _isubsequent tasks, w _i,kexpression task t _iat resource r _kon processing expenditure, R={r ₁, r ₂, r _mthe set of expression heterogeneous resource, succ(t _i) expression task t _isubsequent tasks set; C ' _ijexpression task t _ito task t _jpractical communication expense, r _k=pred_R (t _i) expression task t _jthe resource r choosing _kwith task t _ipre-service resource identical; J=1,2,3 ..., n; K=1,2 ... m;

(1.3) if there is identical rank _sp(t _i) task, to these identical rank _sp(t _i) the rank that calculates according to step (1.2) of task _up(t _i), by rank _up(t _i) descending sort, upgrade scheduling queue;

(1.4) until there is not identical rank _sp(t _i) task, task priority dispatching queue has built;

Above-mentioned j=1,2 ... n; I=1,2 ... n, n represents the sum of task; K=1,2 ... m, m represents the sum of heterogeneous resource;

(2) resource selection of task: according to the priority scheduling queue of the task of above-mentioned structure, the processing resource to task is selected.

2. the real-time task scheduling method based on double priority level according to claim 1, it is characterized in that, in step (1.1), if there are multiple entrance tasks in set of tasks, need to build a virtual entrance task, this virtual entrance task is 0 to the communication overhead between each entrance task; In step (1.2), if there are multiple export tasks in set of tasks, need to build a virtual export task, this virtual export task is 0 to the communication overhead between each export task.

3. the real-time task scheduling method based on double priority level according to claim 1, is characterized in that, for each task, only having after its predecessor task completes and could carry out.

4. the real-time task scheduling method based on double priority level according to claim 1, is characterized in that, the communication overhead of carrying out in same resource of task is 0.