Deadline constraint interleaving workflow scheduling method based on heterogeneous computing system

Abstract

The invention provides a deadline constraint interleaving workflow scheduling method based on a heterogeneous computing system, and the method comprises the steps: obtaining unexecuted tasks on the basis of meeting the topological sorting of each workflow, and calculating the dynamic rising level and sub deadline of each task; calculating the task urgency degree of each task based on the dynamic rising hierarchy and the sub deadline of the task; performing priority ranking on all the ready tasks according to the task urgency to form an ordered list; and for each task in the ordered list, calculating the earliest start time of the task on all candidate calculation units as reference time, converting the end time point of the task on the calculation units into time length, selecting the calculation unit with the lowest energy consumption calculated according to the time length, and distributing the task to the calculation unit. According to the deadline constraint interleaving workflow scheduling method based on the heterogeneous computing system, the problem that an existing workflow scheduling method cannot well balance performance and energy consumption is solved.

Description

Technical Field

[0001] The present invention belongs to the field of workflow scheduling, and in particular relates to a deadline-constrained interleaved workflow scheduling method based on a heterogeneous computing system. Background Art

[0002] System on Chip (SoC) is considered as the backbone of many modern applications, such as medical diagnosis, self-driving cars and avionics. As a complete system, SoC is an integrated circuit containing multiple electronic components, such as processing elements (PE), memory, various interfaces and network communication modules. In view of the shortcomings of homogeneous computing in dealing with such complex applications, people began to think about and explore a new computing model. As a result, heterogeneous computing system (HCS) came into being. Heterogeneous computing system aims to integrate a group of computing elements (PE) with different structures together. In this way, it is expected to better meet the needs of solving a wider range of application problems, thereby breaking through the difficulties faced by homogeneous computing in dealing with complex problems and opening up a new path for the development of computing. HCS combines the advantages of different types of PE (such as general-purpose processors and hardware accelerators), thus providing performance and energy efficiency that homogeneous processing systems cannot achieve.

[0003] To represent applications deployed on SoC platforms, workflow models are widely used and abstracted as directed acyclic graphs (DAGs), where nodes represent tasks and edges represent dependencies between them. Workflow scheduling aims to assign each task in the workflow to a certain time slice of a PE to execute in order to meet the performance criteria specified by the user.

[0004] However, among the existing multi-workflow scheduling methods, some assume that workflows will arrive regularly, and some assume that workflows will arrive sporadically, while there are relatively few technologies for random arrival of workflows, which is the most common scenario in reality. In addition, in addition to the fact that the research direction of existing technologies does not fully fit the practical scenarios, some technologies aim to minimize the execution time of each workflow. However, this does not guarantee that each workflow can be met before the deadline, which leads to extremely high deadline miss rate (DMR) and overall tardiness (OT) outside the deadline. For time-critical applications, HCS (Heterogeneous Computing System) should respond to external stimuli within a specified time limit, which is called the deadline. For example, a key QoS indicator of video streaming is to minimize the deadline miss rate (DMR) of transcoding tasks. In addition, in another part of the technology that pursues meeting the deadline, the optimization of energy consumption (EC) is ignored. Performance and energy consumption are two opposite directions. The blind pursuit of performance will inevitably lead to high energy consumption.

[0005] Therefore, how to implement deadline-constrained interleaved workflow scheduling based on heterogeneous computing systems to better balance performance and energy consumption is an urgent problem to be solved in this field. Summary of the invention

[0006] The purpose of the present invention is to provide a deadline-constrained interleaved workflow scheduling method, system and device based on heterogeneous computing systems in view of the defects of the prior art. The present invention is based on a deadline-constrained interleaved workflow scheduling method for heterogeneous computing systems, calculates the task urgency of each task according to the dynamic ascending level and sub-deadline of the task, and prioritizes the tasks in multiple workflows based on the task urgency, and schedules them according to a list-based strategy to minimize the deadline miss rate DMR and the overall delay OT, while taking into account the energy consumption EC, thereby fundamentally solving the problem that the existing workflow scheduling method cannot balance performance and energy consumption well.

[0007] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0008] The present invention proposes a deadline-constrained interleaved workflow scheduling method based on a heterogeneous computing system, which is characterized by comprising the steps of:

[0009] S1. On the basis of satisfying the topological sorting of each workflow, obtain the unexecuted tasks and calculate the dynamic ascending level and sub-deadline of each task;

[0010] S2, calculate the task urgency of each task based on the dynamic ascending level and sub-deadline of the task;

[0011] S3, prioritizing all ready tasks according to the urgency of the tasks to form an ordered list;

[0012] S4. For each task in the ordered list Calculate the earliest start time est of the task on all candidate computing units as the benchmark time, and use the time on the computing unit as the benchmark time. The end time point is converted into a time length, and a computing unit with the lowest energy consumption calculated according to the time length is selected, specifically including:

[0013] Traverse all candidate computing units, the current computing unit is r m ; If the task has not expired, Where T k For workflow G k Arrival time, For the task The sub-deadline time, ψ is the current time, then the combined indicator where ft m for In r m The earliest time that the execution is completed, express In r m The execution time, For the kth workflow G k The i-th task of is the mth computing unit r m otherwise, set it to Finally, the task Assigned to the computing unit with the smallest J value, wherein the computing unit includes a processor and an accelerator, and the computing units are interconnected.

[0014] Furthermore, the calculation of the dynamic ascending level and sub-deadline of each task in step S1 specifically includes:

[0015] S11. For the newly arrived workflow G v and the old unfinished workflow G u ,Different methods are used to calculate the dynamic ascending level of tasks;

[0016] S12. Calculate the sub-deadline of each task.

[0017] Furthermore, step S11 specifically includes:

[0018] For the new workflow Gv , no task starts to execute, get its task by the following way Dynamic ascending levels:

[0019]

[0020] in is from arrive The average communication time, is the average communication speed, It's a task The average computation time across all computational units, For the task Dynamic ascending level; by v The beginning node of Traverse the DAG downwards and recursively calculate the workflow G v Dynamic ascending hierarchy of all tasks in;

[0021] Wherein, the DAG is a weighted directed acyclic graph DAG of the workflow, DAG = (N, E), where node n in N i ∈N represents the i-th task, and the directed edge e in E i,j ∈E represents the priority dependency between the i-th and j-th tasks. Two virtual tasks n with zero execution time are entry and n exit Add to the beginning and end of the workflow respectively.

[0022] Furthermore, step S11 specifically includes: for the currently executing workflow G u , calculate its task Dynamic ascending level The specific steps are:

[0023] When calculating the dynamic ascending level, the uncertain task Will it be reassigned to the previously selected computational unit? The consistency of this assignment is estimated as a probability,

[0024]

[0025] in, Indicates keeping the original distribution probability, is the transmission time between the specified source and target computing units, is from arrive The average communication time, is the average communication speed, For the task The average computation time across all computational units, For the task The actual completion time, ψ is the current time, For the task In the mth computing unit r m Execution time; From the task arrive The average communication time of , whose source computing unit is determined, but the target computing unit is uncertain;

[0026] If the distribution consistency holds, then from ψ to completion The time required is Otherwise it is estimated to be

[0027] When this consistency is established, from the completed predecessor node arrive The data arrival time is Otherwise

[0028] Based on Combining the arrival times of two data to estimate The point in time when the node ends receiving data from its completed parent node, i.e.,

[0029]

[0030] The earliest start time is the time when all data are received from its parent node;

[0031] By analyzing the predecessor nodes in all states, the measurement from ψ to The duration of the start time is calculated Dynamic top layer Specifically:

[0032]

[0033] Furthermore, step S12, calculating the sub-deadline of each task, specifically includes:

[0034] node Sub-deadline The calculation is as follows:

[0035]

[0036] in, for The dynamic ascending level indicates that from Start time Duration of the end time; Dk For workflow G k The deadline, T k For workflow G k The arrival time of , ψ is the current time, for The dynamic ascending level, For the task Average computation time across all compute units.

[0037] Further, step S2, calculating the task urgency of each task based on the dynamic ascending level and the sub-deadline of the task, specifically includes:

[0038] The urgency of the task is:

[0039]

[0040] Among them, ω is the weight parameter, and its value is [0,1]; the lower Value represents task More urgent, should be assigned to the computing unit as early as possible.

[0041] The present invention also proposes a deadline-constrained interleaved workflow scheduling system based on a heterogeneous computing system, characterized in that the workflow scheduling system executes the deadline-constrained interleaved workflow scheduling method based on a heterogeneous computing system, including: a dynamic ascending level and sub-deadline time calculation module, a task urgency calculation module, an ordered list generation module, and a task scheduling module;

[0042] The dynamic ascending level and sub-deadline calculation module obtains unexecuted tasks and calculates the dynamic ascending level and sub-deadline of each task on the basis of satisfying the topological sorting of each workflow;

[0043] The task urgency calculation module calculates the task urgency of each task based on the dynamic ascending level and sub-deadline of the task;

[0044] An ordered list generation module is used to prioritize all ready tasks according to the urgency of the tasks to form an ordered list;

[0045] The task scheduling module, for each task in the ordered list Calculate the earliest start time est of the task on all candidate computing units as the benchmark time, and use the time on the computing unit as the benchmark time. The end time point is converted into a time length, and the computing unit with the lowest energy consumption calculated according to the time length is selected

[0046] The present invention also proposes a scheduler, which executes the deadline constraint interleaved workflow scheduling method based on heterogeneous computing systems.

[0047] Compared with the prior art, it has the following beneficial effects:

[0048] The deadline-constrained interleaved workflow scheduling method based on heterogeneous computing systems of the present invention calculates the task urgency of each task according to the dynamic ascending level and sub-deadline of the task for the interleaved workflow, prioritizes the tasks in multiple workflows based on the task urgency, and schedules them according to a list-based strategy. A list scheduling heuristic method based on task urgency is used to schedule each newly injected workflow and the unfinished tasks of the previous workflow. The method implements the scheduling of interleaved multiple workflows in a SoC heterogeneous computing system with interconnected computing units, and introduces deadline constraints to minimize the deadline miss rate DMR and the overall delay OT; and while optimizing performance, energy consumption EC is fully considered to better balance performance and energy consumption in workflow scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0050] Figure 1 A schematic diagram of a scheduling framework for an interleaved workflow provided in an embodiment of the present invention.

[0051] Figure 2 The radar correlator workflow is executed on a SoC platform provided in an embodiment of the present invention, wherein (a) is a DAG of the workflow, and (b) is a calculation time matrix.

[0052] Figure 3 A schematic diagram of a deadline-constrained interleaved workflow scheduling method based on a heterogeneous computing system provided in an embodiment of the present invention.

[0053] Figure 4 A schematic diagram of the structure of a workflow application provided by an embodiment of the present invention;

[0054] Figure 5 A schematic diagram of EC, DMR and OT values ​​of λ under different cut-off time settings for the energy sensing method provided in an embodiment of the present invention, Figure 5 (a) is the EC under loose settings, Figure 5 (b) is the EC under medium setting, Figure 5 (c) is the EC under compact setting, Figure 5 (d) is the DMR under relaxed settings, Figure 5 (e) is the DMR under medium setting, Figure 5(f) is the DMR in a compact setting, Figure 5 (g) is the OT under loose settings, Figure 5 (h) is the OT at medium setting, Figure 5 (i) is the OT in a compact setting.

[0055] Figure 6 A schematic diagram of a deadline-constrained interleaved workflow scheduling system based on a heterogeneous computing system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0058] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0059] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0060] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0061] In the present invention, the relevant parameters are described as follows:

[0062]

[0063] Due to the huge performance and EC heterogeneity of PEs in today’s SoCs (e.g., PEs based on large .LITTLE architectures), it is more critical to select the right PE to perform specific performance and EC optimization tasks. Hereafter, to develop an energy-aware variant of the multi-workflow scheduling algorithm ULS, ULSE, our main goal is to exploit the heterogeneity in SoCs.

[0064] For real-time SoC, workflow applications can be injected at any time, that is, a new workflow may arrive before the previous workflow is completed. Therefore, SoC needs to execute tasks from different workflows simultaneously. To this end, the present invention proposes a scheduling framework for SoC platform, such as Figure 1 As shown, to cope with the arrival and execution intertwined workflows.

[0065] When a new workflow arrives, all its tasks are injected into the system's pool of unfinished tasks. In addition, there may still be tasks from previous workflows in the pool. For any workflow, tasks are ready when all predecessor tasks have been completed. Otherwise, they will wait in the waiting pool.

[0066] The scheduler then calls the scheduling algorithm to schedule all tasks in the pool and assigns a i Determine a four-tuple sched i = <n i ,st i ,ft i ,r m >, which means PEr m Start timest i and end time ft i The time slices between are allocated to n i Although tasks from previous workflows in the pool have already been scheduled, the scheduler can make smarter decisions by rescheduling them along with new workflows to better utilize SoC resources and enable more workflows to complete on time.

[0067] When ready tasks become available on a specific PE, they are immediately deployed to the PE for execution. The information of completed tasks is fed back to the scheduler to distinguish between ready tasks and waiting tasks.

[0068] Complex applications running on SoC platforms may contain many computing tasks. The workflow model can be used to represent SoC applications through a weighted directed acyclic graph DAG = (N, E), where node n in N i ∈N represents the i-th task, and the directed edge e in E i,j ∈E represents the priority dependency between the i-th and j-th tasks. Each task n i are inseparable, each priority dependency e i,j Represents task n j Only after completing n i In addition, if e i,j With data communication, then n j Only after receiving n iThe source and target of an edge are called the parent task and child task respectively. i The immediate predecessor set is *n i Indicates that all direct successor sets are composed of n i * indicates. In order to generalize to a workflow with one entry and one exit, two virtual tasks n with zero execution time are added. entry and n exit Add to the beginning and end of the workflow respectively.

[0069] When a workflow application is deployed on a SoC platform, each of its tasks can be executed on a different PE, and the performance of each task on the PE can be obtained through analytical modeling, empirical modeling, or historical data. m Previous task i The required computation time w i,m . Figure 2 (a) shows a radar workflow application called RadarCorrelator. In addition to the two virtual tasks n entry and n exit In addition, it also consists of 6 tasks. The edge weights in the figure represent the amount of data transferred between tasks. Assume that it is deployed on a SoC platform consisting of Odroid-XU3 and 4 FFT accelerators. According to data analysis, the execution time of each task on each PE is as follows: Figure 2 (b) shows that the execution time required for tasks on different PEs varies greatly. For example, the execution time of n5 on A7 is 2.5 times that of A15 and 5 times that of FFT accelerator. In summary, the heterogeneity of performance and energy consumption of PEs in SoC is very large.

[0070] For the kth workflow G k , let T k Its arrival time, D k is the deadline. That is, G k Should be in T k +D k Completed before. Assume that for G k , whose mission and Deployed in PEr m and r n There is a priority dependency between them. The amount of data is Transmission time Depends on r m and r n Between and communication speed b m,n When r m and r nWhen referencing the same PE, it becomes zero, as shown:

[0071]

[0072] When m When its leading data are ready, Can be in r m Start execution on Yes m Can be used to execute Therefore, r m superior The start and end time of can be obtained as follows, where express In PEr m Execution time on:

[0073]

[0074] For real-time SoCs, workflows may arrive unpredictably, and the scheduler needs to complete as many workflows as possible quickly. When the deadline of some workflows is violated, their overall delay should be as small as possible. Specifically, the value of DMRρ is defined as:

[0075]

[0076] in is the set of all workflows injected into the SoC platform. If the condition is met, the indicator function I(·) returns 1, otherwise it returns 0.

[0077] The value of OTη can be calculated as,

[0078]

[0079] Energy consumption is also critical for SoC platforms. Let ω be the EC value of the SoC platform processing all workflows. It can be obtained as follows:

[0080]

[0081] in, is the energy consumption of the mth PE at time point t, and Execute all workflows respectively The start time and end time of the

[0082] Therefore, our problem can be formulated as how the scheduler can perform dynamic scheduling to minimize DMR, OT, and EC.

[0083] In order to solve the above optimization problem, the present invention proposes an energy-aware scheduling heuristic ULSE. When a new workflow arrives, there may be one, multiple, or no currently executing workflows in the HCS. Without loss of generality, we assume that when a new workflow G v When reaching the current time point ψ, there is a workflow G u Executing, where ψ = T v >T u In order to do our best to ensure G u and G v Completed within their respective deadlines, G u Tasks that have not yet been executed in G v Reschedule the tasks.

[0084] To reduce energy consumption, unlike most two-stage scheduling algorithms, ULSE only processes ready tasks instead of all unfinished tasks to dynamically use more accurate real-time power values ​​of PE and make smarter decisions. Whenever some tasks become ready, they are inserted into the ordered list of ready tasks in ascending order of priority value.

[0085] The present invention proposes a deadline-constrained interleaved workflow scheduling method based on heterogeneous computing systems. Figure 3 As shown, the deadline-constrained interleaved workflow scheduling method based on heterogeneous computing systems includes the following steps S1 to S4.

[0086] S1. On the basis of satisfying the topological sorting of each workflow, obtain the unexecuted tasks and calculate the dynamic ascending level and sub-deadline of each task.

[0087] Task i The ascending level is widely used for static measurement from the start time of the entry task to n i The duration of the earliest start time. Since in dynamic scenarios, the workflow may already be in progress, we introduce a dynamic ascending level τ i To measure from the current time point ψ to n i The duration of the earliest start time.

[0088] In step S1, the dynamic ascending level and sub-deadline of each task are calculated, which specifically includes:

[0089] S11. For the newly arrived workflow G v and the old unfinished workflow G u ,different methods are used to calculate the dynamic ascending level of tasks.

[0090] For the new workflow G v , no task starts to execute, get its task by the following way Dynamic ascending levels:

[0091]

[0092] in is from arrive The average communication time, is the average communication speed, It's a task The average computation time across all computational units, For the task Dynamic ascending level; by v The beginning node of Traverse the DAG downwards and recursively calculate the workflow G v Dynamic ascending hierarchy of all tasks in;

[0093] Wherein, the DAG is a weighted directed acyclic graph DAG of the workflow, DAG = (N, E), where node n in N i ∈N represents the i-th task, and the directed edge e in E i,j ∈E represents the priority dependency between the i-th and j-th tasks. Two virtual tasks n with zero execution time are entry and n exit Add to the beginning and end of the workflow respectively.

[0094] For the currently executing G u , calculate its task Dynamic ascending level More complex because it needs to include real-time execution information. u Pending Ready Tasks The parent task can only be Or completed. Let F u G u The set of tasks currently completed in . If The actual completion time (aft) of the completed parent is obtained. Since they may have already or are in the process of contributing to the current task data is being transmitted, so rescheduling may result in data being retransmitted.

[0095] When calculating the dynamic ascending level, the uncertain task Will it be reassigned to the previously selected computational unit? The consistency of this assignment is estimated as a probability,

[0096]

[0097] in, Indicates keeping the original distribution probability, is the transmission time between the specified source and target computing units, is from arrive The average communication time, is the average communication speed, For the task The average computation time across all computational units, For the task The actual completion time, ψ is the current time, For the task In the mth computing unit r m Execution time; From the task arrive The average communication time of a communication whose source computing unit is determined and the target computing unit is uncertain.

[0098] Specifically, when A larger value, that is, closer to 1, indicates that allocation consistency is more likely to be maintained. On the contrary, it means that the overhead of reallocation is tolerable in overall optimization, so reallocation is more likely to be maintained.

[0099] If the distribution consistency holds, then from ψ to completion The time required is Otherwise it is estimated to be When the ratio between them is small, A larger value of , that is, closer to 1, indicates that the distribution consistency is more likely to be maintained.

[0100] When this consistency is established, from the completed predecessor node arrive The data arrival time is Otherwise

[0101] Based on Combining the arrival times of two data to estimate The point in time when the node ends receiving data from its completed parent node, i.e.,

[0102]

[0103] The earliest start time is the time when all data are received from its parent node;

[0104] By analyzing the predecessor nodes in all states, the measurement from ψ to The duration of the start time is calculated Dynamic top layer Specifically:

[0105]

[0106] By traversing the workflow graph downward, we can recursively calculate G u The properties of each uncompleted task in .

[0107] S12. Calculate the sub-deadline of each task.

[0108] Step S12: Calculate the sub-deadline of each task, specifically including:

[0109] node Sub-deadline The calculation is as follows:

[0110]

[0111] in, for The dynamic ascending level indicates that from Start time Duration of the end time; D k For workflow G k The deadline, T k For workflow G k The arrival time of , ψ is the current time, for The dynamic ascending level, For the task Average computation time across all compute units.

[0112] S2. Calculate the task urgency of each task based on the task's dynamic ascending level and sub-deadline.

[0113] The urgency of a task is reflected in its earliest start time and latest finish time, which are measured by the dynamic top-level τ and the sub-deadline d, respectively. The former is crucial to make more tasks in the workflow ready, while the latter directly determines whether the workflow deadline can be met.

[0114] Step S2: Calculate the task urgency of each task based on the dynamic ascending level and sub-deadline of the task, specifically including:

[0115] The urgency of the task is:

[0116]

[0117] Among them, ω is the weight parameter, and its value is [0,1]; the lower Value represents task More urgent, should be assigned to the computing unit as early as possible.

[0118] For a given workflow D k Any two unfinished tasks in and Always less than And in the task list obtained by sorting tasks according to urgency values, all priority dependencies between tasks in the workflow diagram are respected, which is a prerequisite for list scheduling.

[0119] S3. Prioritize all ready tasks according to the task urgency to form an ordered list.

[0120] This technology adopts the classic two-stage list scheduling. The first stage prioritizes the tasks according to the sorting factor, and the second stage allocates computing units to the selected tasks in order.

[0121] In the priority sorting stage, ULSE sequentially sorts all ready tasks in ascending order of μ, which in turn serves as the input to the computational unit selection stage.

[0122] In the computing unit selection stage, the present invention considers the energy consumption and completion time of the task to balance these two goals.

[0123] S4. For each task in the ordered list Calculate the earliest start time est of the task on all candidate computing units as the benchmark time, and use the The end time point is converted into a time length;

[0124] The st of any node in the set is associated with two aspects: one is related to the completion of the predecessor execution and the end of data transfer; the other is related to the preparation time of the newly assigned PE. The former is further related to the selection of the computing unit PE.

[0125] Next, a computing unit with the lowest energy consumption calculated based on the time length is selected, including:

[0126] Traverse all candidate computing units, the current computing unit is r m ; If the task has not expired, Where T k For workflow G k Arrival time, For the task The sub-deadline time, ψ is the current time, then the combined indicator where ft m for In r m The earliest time that the execution is completed, express In rm The execution time, For the kth workflow G k The i-th task, p m is the mth computing unit r m Otherwise, set it to to give the completion time a greater weight in the combined evaluation; finally, Assigned to the computing unit with the smallest J value, wherein the computing unit includes a processor and an accelerator, and the computing units are interconnected; data transmission is performed point-to-point between processors, between processors and accelerators, and between accelerators.

[0127] The present invention proposes a collaborative scheduling model ULSE for deadline-constrained interleaved workflows (including dynamically arriving workflows) executed on heterogeneous SoCs to minimize DMR, OT, and EC. In order to minimize DMR and OT, the present invention uses the task attribute of task urgency to prioritize tasks in multiple workflows and schedule them according to a list-based strategy. In addition, the present invention also considers energy consumption, and the workflow scheduling method can better balance performance and EC.

[0128] Furthermore, the proposed approach is evaluated on a simulated SoC, which has been carefully validated on the Odroid-XU3 platform. Using a dynamically interleaved workload mix, the experimental results validate the superiority of the proposed workflow scheduling approach.

[0129] DS3 is an open-source, advanced SoC simulator specifically designed to evaluate scheduling and power management algorithms on different hardware configurations. It has been validated against real SoC platforms, with the difference between simulation results and reference real measurements being less than 3% on average. We implemented the proposed scheduling framework in DS3 and conducted evaluation experiments on a PC equipped with a 12th-generation Intel Core i7-12700H and Windows 11.

[0130] The following heterogeneous SoC platform in DS3 is used for evaluation: The Odroid-XU3 board consists of 4 ARM Cortex-A7 cores and 4 ARM Cortex-A15 cores, and is also equipped with two matrix multiplication accelerators, four FFT accelerators, and two Viterbi decoder accelerators. The workload mix used for evaluation consists of five workflow applications available in DS3, namely Radar Correlator, Temporal Mitigation, WiFi Transmitter, WiFi Receiver, and Custom Workflow. Their structure is as follows Figure 4 shown.

[0131] Workflows are continuously injected into the SoC framework, and each injection uses one of the five workflows at random. It is assumed that the injection interval between two consecutive workflows follows an exponential distribution Exp(λ). More precisely, the parameter λ represents the expected number of workflow arrivals in one second, and 1 / λ represents the expected arrival interval between two consecutive workflows. Larger values ​​of λ indicate higher workloads in the considered SoC platform.

[0132] To set a deadline for each workflow on the SoC platform, we first schedule it using HEFT and execute it to record its completion time (denoted as M). As shown in the following table, we introduce three types of deadlines based on M: tight, medium, and loose. For example, for a tight workflow, the deadline of the workflow is set to 2 times its completion time with a probability of 80%, and 3 times with a probability of 20%.

[0133] Three types of settings

[0134] type 2*M 3*M compact 80％ 20％ medium 50％ 50％ Loose 20％ 80％

[0135] Comparison Techniques The following state-of-the-art methods are used to compare the proposed method: HEFT RT 、HEFT EDP-LB and PEFT RT HEFT RT 、HEFT EDP-LB There are two multi-workflow scheduling methods: the former aims to minimize the completion time of each workflow; while the latter considers the load balancing of energy consumption and execution time. RT HEFT is a variant of PEFT, a single workflow scheduling method for multiple workflows. EDP-LB Both are EC optimization technologies, so DMR and OT perform poorly; HEFT RT and PEFT RT It performs well in DMR and OT, but poorly in EC.

[0136] A total of 1000 workflows are randomly injected for experimental evaluation. To illustrate the performance comparison between scheduling algorithms, we use the following metrics: Deadline Miss Rate (DMR), Overall Latency (OT), and Energy Consumption (EC).

[0137] The conclusions of the experiment are expected to be: 1. The ULSE of this technology is the lowest on EC, followed by HEFTEDP-LB, and the worst are HEFTRT and PEFTRT. 2. The ULSE on DMR and OT is comparable to HEFTRT and PEFTRT, and is better than HEFTEDP-LB.

[0138] We compare the proposed ULSE with the existing energy-aware multi-workflow scheduling method HEFTEDP-LB, increasing λ from 32 to 50 in three deadline settings. We also use two best-effort HEFTRT and PEFTRT as the baselines for this experiment. The experimental results are shown in Figure 5 shown.

[0139] The energy consumption results of each method are shown in Figure 5 (a)-(c). As λ increases, the EC required for each method decreases. This is because in the case of each λ, 1000 workflows are executed, and when λ is larger, the total time to execute these 1000 workflows decreases, resulting in a lower EC value. The two energy-aware methods consume much less energy than the two best-effort methods. Compared with ULS, ULSE reduces the energy consumption of executing 1000 workflows by an average of 12%. When λ is smaller, the reduction is more significant, for example, when λ = 32 and the deadline type is loose, the reduction is 21%. Compared with HEFT EDP-LB In comparison, the energy consumption of ULSE is lower in all cases. For looser deadline settings, the EC of ULSE is generally slightly lower. This is because ULSE tends to use more energy-efficient PEs when the deadline is less stringent. For example, when λ is 36, the EC of ULSE is 335.9 J under loose deadlines and rises to 337.8 J under tight deadlines. RT HEFT EDP-LB Workflow deadlines are not considered, so their EC values ​​in the three deadline settings are generally the same.

[0140] The corresponding DMR and OT values ​​for each method are shown in Figure 5 (d)-(i). For all λ and deadline settings, ULSE has significantly better DMR and OT values ​​than HEFT. EDP-LB Specifically, with HEFT EDP-LB In comparison, ULSE reduces the DMR and OT values ​​by 15.3% and 45.8% on average, respectively. Even under multiple test points, ULSE is even better than the best-effort method HEFT. RT and PEFT RT Achieved better performance. With the shortening of deadline and the increasing of λ value, the DMR and OT values ​​of each method generally increase.

[0141] In summary, compared with the existing energy-aware method HEFT EDP-LB Compared with HEFT, which pursues performance, the workflow scheduling method ULSE of the present invention consumes less energy and performs better in DMR and OT. RT and PEFT RTIn comparison, ULSE consumes less energy and is comparable to DMR and OT at multiple test points.

[0142] Figure 6 The embodiment of the present invention provides a deadline-constrained interleaved workflow scheduling system based on a heterogeneous computing system. Figure 6 As shown, the deadline-constrained interleaved workflow scheduling system based on heterogeneous computing systems includes a dynamic ascending level and sub-deadline calculation module, a task urgency calculation module, an ordered list generation module, and a task scheduling module.

[0143] The dynamic ascending level and sub-deadline calculation module obtains unexecuted tasks and calculates the dynamic ascending level and sub-deadline of each task on the basis of satisfying the topological sorting of each workflow;

[0144] The task urgency calculation module calculates the task urgency of each task based on the dynamic ascending level and sub-deadline of the task;

[0145] An ordered list generation module is used to prioritize all ready tasks according to the urgency of the tasks to form an ordered list;

[0146] The task scheduling module, for each task in the ordered list Calculate the earliest start time est of the task on all candidate computing units as the benchmark time, and use the The end time point is converted into a time length, and a computing unit with the lowest energy consumption calculated according to the time length is selected.

[0147] The present invention also proposes a scheduler, which executes the deadline constraint interleaved workflow scheduling method based on heterogeneous computing systems.

[0148] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0149] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of each unit is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0150] The steps in the method of the embodiment of the present invention can be adjusted in order, combined and deleted according to actual needs. The units in the device of the embodiment of the present invention can be combined, divided and deleted according to actual needs. In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0151] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, terminal, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention.

[0152] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A deadline-constrained interleaved workflow scheduling method based on heterogeneous computing systems, characterized in that: Includes steps: S1. On the basis of satisfying the topological sorting of each workflow, obtain the unexecuted tasks and calculate the dynamic ascending level and sub-deadline of each task; S2, calculate the task urgency of each task based on the dynamic ascending level and sub-deadline of the task; S3, prioritizing all ready tasks according to the urgency of the tasks to form an ordered list; S4. For each task in the ordered list Calculate the earliest start time est of the task on all candidate computing units as the benchmark time, and use the time on the computing unit as the benchmark time. The end time point is converted into a time length, and a computing unit with the lowest energy consumption calculated according to the time length is selected, specifically including: Traverse all candidate computing units, the current computing unit is r m ; If the task has not expired, Where T k For workflow G k Arrival time, For the task The sub-deadline time, ψ is the current time, then the combined indicator where ft m for In r m The earliest time that the execution is completed, express In r m The execution time, For the kth workflow G k The i-th task of is the mth computing unit r m otherwise, set it to Finally, the task Assigned to the computing unit with the smallest J value; wherein the computing unit includes a processor and an accelerator, and each computing unit is interconnected.

2. The method according to claim 1, characterized in that In step S1, the dynamic ascending level and sub-deadline of each task are calculated, which specifically includes: S11. For the newly arrived workflow G v and the old unfinished workflow G u ,Different methods are used to calculate the dynamic ascending level of tasks; S12. Calculate the sub-deadline of each task.

3. The method according to claim 2, characterized in that Step S11 specifically includes: For the new workflow G v , no task starts to execute, get its task by the following way Dynamic ascending levels: in is from arrive The average communication time, is the average communication speed, It's a task The average computation time across all computational units, For the task Dynamic ascending level; by v The beginning node of Traverse the DAG downwards and recursively calculate the workflow G v Dynamic ascending hierarchy of all tasks in; Wherein, the DAG is a weighted directed acyclic graph DAG of the workflow, DAG = (N, E), where node n in N i ∈N represents the i-th task, and the directed edge e in E i,j ∈E represents the priority dependency between the i-th and j-th tasks. Two virtual tasks n with zero execution time are entry and n exit Add to the beginning and end of the workflow respectively.

4. The method according to claim 2, characterized in that: Step S11 specifically includes: for the currently executing workflow G u , calculate its task Dynamic ascending level The specific steps are as follows: when calculating the dynamic ascending level, the uncertain task Will it be reassigned to the previously selected computational unit? The consistency of this assignment is estimated as a probability, in, Indicates keeping the original distribution probability, is the transmission time between the specified source and target computing units, is from arrive The average communication time, is the average communication speed, For the task The average computation time across all computational units, For the task The actual completion time, ψ is the current time, For the task In the mth computing unit r m Execution time; From the task arrive The average communication time of , whose source computing unit is determined, but the target computing unit is uncertain; If the distribution consistency holds, then from ψ to completion The time required is Otherwise it is estimated to be When this consistency is established, from the completed predecessor node arrive The data arrival time is Otherwise Based on Combining the arrival times of two data to estimate The point in time when the node ends receiving data from its completed parent node, i.e., The earliest start time is the time when all data are received from its parent node; By analyzing the predecessor nodes in all states, the measurement from ψ to The duration of the start time is calculated Dynamic ascending level Specifically:

5. The method according to claim 2, characterized in that: Step S12: Calculate the sub-deadline of each task, specifically including: node Sub-deadline The calculation is as follows: in, for The dynamic ascending level indicates that from Start time is up Duration of the end time; D k For workflow G k The deadline, T k For workflow G k The arrival time of , ψ is the current time, for The dynamic ascending level, For the task Average computation time across all compute units.

6. The method according to claim 1, characterized in that Step S2: Calculate the task urgency of each task based on the dynamic ascending level and sub-deadline of the task, specifically including: The urgency of the task is: Among them, ω is the weight parameter, and its value is [0,1]; the lower Value represents task More urgent, should be assigned to the computing unit as early as possible.

7. A deadline-constrained interleaved workflow scheduling system based on heterogeneous computing systems, characterized in that: The workflow scheduling system executes the deadline constraint interleaved workflow scheduling method based on heterogeneous computing systems as claimed in claim 1, comprising: a dynamic ascending level and sub-deadline calculation module, a task urgency calculation module, an ordered list generation module, and a task scheduling module; The dynamic ascending level and sub-deadline calculation module obtains unexecuted tasks and calculates the dynamic ascending level and sub-deadline of each task on the basis of satisfying the topological sorting of each workflow; The task urgency calculation module calculates the task urgency of each task based on the dynamic ascending level and sub-deadline of the task; An ordered list generation module is used to prioritize all ready tasks according to the urgency of the tasks to form an ordered list; The task scheduling module, for each task in the ordered list Calculate the earliest start time est of the task on all candidate computing units as the benchmark time, and use the The end time point is converted into a time length, and a computing unit with the lowest energy consumption calculated according to the time length is selected.

8. A scheduler, wherein the scheduler executes the deadline-constrained interleaved workflow scheduling method based on heterogeneous computing systems as described in claim 1.

Images