CN107861606A - A kind of heterogeneous polynuclear power cap method by coordinating DVFS and duty mapping - Google Patents

Abstract

The invention discloses a heterogeneous multi-core power capping method by coordinating DVFS and task mapping. Firstly, for heterogeneous systems, the scripts that can measure the power consumption of computing nodes, CPU power consumption and GPU power consumption respectively after program execution is completed, and then modify The selected parallel test benchmark program is used to obtain the execution time of different kernel functions; then, at different frequencies of the CPU and GPU, run the application program only on the CPU and GPU respectively to obtain detailed running information, including the total execution time, each Kernel function execution time, computing node power consumption, CPU power consumption and GPU power consumption; based on the running information, design a prediction model, including the prediction execution time model and power consumption model; finally, based on the prediction model, get different CPU frequencies and GPU frequencies Fill in the configuration table with the system power consumption and execution time under the task allocation scheme, and find the optimal configuration scheme according to the improved greedy algorithm. By adopting the invention, the system performance can be improved and the system power consumption budget can be limited at the same time.

Description

A Power Capping Method for Heterogeneous Multi-Cores by Coordinating DVFS and Task Mapping

technical field

The invention belongs to the field of computer architecture, and in particular relates to realizing a heterogeneous multi-core power capping method by coordinating DVFS and task mapping.

Background technique

After continuous research and development in recent years, the advanced architecture represented by multi-core processors has gradually replaced single-core processors as the main way to improve processor performance. Compared with homogeneous multi-core processors, heterogeneous multi-core platforms can achieve better performance. Power capping is a technique that limits the power consumption of a heterogeneous system to a predetermined level. Power consumption and heat dissipation limit the improvement of heterogeneous multi-core performance. Modern processors are built so that they can suffer from a certain level of power consumption, requiring systems that can achieve a processor power cap. The most common power budgeting techniques today rely on hardware components operating at different frequencies and thus have different power consumption, the main idea is to use dynamic voltage frequency scaling (DVFS). While using DVFS to limit the power consumption of heterogeneous systems, there will be load imbalance between CPU and GPU. By decomposing a parallel program into tasks that can be executed simultaneously, and mapping each task to the most suitable processor, the computing power of the system can be fully utilized to improve system performance, but this mapping scheme usually does not take system power consumption into consideration. This paper proposes a scheme that combines DVFS and task mapping to improve system performance while limiting system power consumption budget.

Contents of the invention

The present invention proposes a heterogeneous multi-core power capping method by coordinating DVFS and task mapping to improve system performance while limiting system power consumption budget. Firstly, for heterogeneous systems, a method that can measure the power consumption of computing nodes after program execution is completed , the scripts of CPU power consumption and GPU power consumption, and then modify the selected parallel test benchmark program to obtain the execution time of different kernel functions. Then, at different frequencies of the CPU and GPU, run the application only on the CPU and GPU respectively, and obtain detailed running information, including the total execution time, the execution time of each kernel function, the power consumption of computing nodes, CPU power consumption and GPU power consumption . Based on the running information, a predictive model is designed, including predictive execution time model and power consumption model. Finally, based on the prediction model, the system power consumption and execution time under different CPU frequencies, GPU frequencies and task allocation schemes are obtained and filled in the configuration table. According to the improved greedy algorithm, the optimal configuration scheme (CPU frequency, GPU frequency, task mapping table) is found.

In order to achieve the above object, the present invention adopts the following technical solutions.

A heterogeneous multi-core power capping method by coordinating DVFS and task mapping, combining DVFS and task mapping to limit system power consumption under budget power consumption while pursuing system performance; comprising the following steps:

Step 1, implement measuring the total execution time of the application program, the power consumption of the CPU and the power consumption of the GPU after the execution of the application program is completed.

Step 2, modify the selected parallel test benchmark program, and obtain the execution time of each kernel function in the program.

Step 3. Execute the application program only on the CPU or GPU respectively, set different CPU (GPU) optional frequencies, and obtain detailed running information, including the total execution time of the program, the execution time of each kernel function in the program, and the total power of the system. Power consumption, CPU power consumption and GPU power consumption.

Step 4, designing a prediction model to predict the power consumption and execution time of different task mapping schemes at different CPU and GPU frequencies. The input of the prediction model is the CPU frequency, GPU frequency and task mapping scheme. The task mapping involved in this article refers to mapping each kernel function as a whole to the CPU or GPU, rather than assigning the kernel function to the CPU according to a certain ratio. Execute simultaneously on the GPU. The output of the predictive model is the predicted total program execution time and the predicted system power consumption. This predictive model includes an execution time model and a power consumption model.

Step 4.1, Execution Time Model

The total execution time of the application program can be obtained according to the execution time of each core in the program and the corresponding data transmission time. The total execution time of the program is expressed by the formula ①.

Among them, f _cpu and f _gpu represent the CPU frequency and GPU frequency respectively, and T _i (f _cpu , f _gpu ) represents the execution time of the i-th kernel function and the required data transmission time, expressed by formula ②.

The first part of the formula represents the execution time, and the second part represents the data transfer time. The execution time of the kernel function has been obtained in step 2. H2D and D2H denote the data transfer cost from host to device and the transfer cost from device to host, respectively. Required is equal to 1 or 0, indicating whether the core k is required or the data d is required. The data may already be on the device and does not need to be transmitted, so use OnDevice to indicate whether the data is on the device. Use size to represent the data size.

Step 4.2, Power Consumption Model

System power consumption may be represented by three parts, which are idle power consumption, CPU power consumption P _cpu and GPU power consumption P _gpu . System power consumption is represented by formula ③.

P＝P _idle (f _cpu ，f _gpu )+P _cpu (f _cpu ，f _gpu )+P _gpu (f _cpu ，f _gpu ) ③

Among them, P _idle represents idle power consumption, which is related to CPU frequency and GPU frequency, but has nothing to do with task mapping. It can be obtained by subtracting the power consumption on the CPU from the total power consumption obtained by executing the application program only on the CPU, and obtained by formula ④.

Respectively represent the total power consumption of the system, CPU power consumption and GPU power consumption obtained under the condition that only the application program is executed on the CPU, and are related to the set frequency.

When the application program is executed, the CPU and GPU are not always executing due to the data association between the kernel functions, so the CPU power consumption and GPU power consumption vary according to the given task mapping scheme, combined with this phenomenon , we assume that power consumption and execution time are proportional to the ratio of total execution time, so CPU and GPU power consumption can be expressed by formula ⑤ and formula ⑥ respectively.

Among them, λ _cpu and λ _gpu represent the CPU and GPU busy time ratio, respectively, Respectively represent the CPU power consumption and the GPU power consumption obtained under the condition that the application program is only executed on the GPU. with estimators representing the maximum value of dynamic power consumption in the CPU and GPU, respectively.

Since _λcpu and _λgpu denote CPU and GPU busy time ratios respectively, they can be obtained from the running information measured in step 3. When the program is executed, the busy time of each device is defined as the sum of the actual execution time of each kernel function. λ _cpu and λ _gpu can be represented by formula ⑦ and formula ⑧, respectively.

t _cpu represents the CPU busy time, t _gpu represents the GPU busy time, and t _t o _tal represents the total execution time of the application.

Step 5: Construct a configuration parameter table based on the prediction model.

According to the execution time model and power consumption model, the execution time and power consumption under different CPU frequencies and GPU frequencies, different task mapping schemes are calculated, and filled in the configuration parameter table.

Step 6, according to the configuration parameter table, use the improved greedy algorithm to search for the optimal parameter set. It is divided into two steps. First, use the greedy algorithm to search for the task mapping scheme with the shortest execution time under the given CPU frequency and GPU frequency. According to the task mapping, CPU frequency, GPU frequency and power model, the system prediction under this parameter configuration is obtained. power consumption. Then, change the CPU frequency and GPU frequency, calculate the predicted power consumption of the system again according to the previous step, and finally obtain the optimal configuration scheme limited to the budgeted power consumption.

Step 6.1, given the CPU frequency and GPU frequency, search for the task mapping scheme with the shortest execution time, and obtain the predicted power consumption of the system according to the power consumption model.

Step 6.2, according to the optional settings of CPU frequency and GPU frequency, change the CPU frequency and GPU frequency, repeat step 6.1, obtain the mapping scheme corresponding to the optimal execution time under this frequency combination, and calculate the system power consumption according to this mapping scheme, Based on the given budget power consumption, the optimal frequency parameter selection and mapping scheme limited under the budget power consumption is obtained.

Compared with the prior art, the present invention has the following advantages:

Combining DVFS and task mapping together realizes limiting system power consumption to budget power consumption while ensuring system performance. Most of the existing system power capping technologies are implemented through dynamic voltage and frequency scaling, because the device frequency in the system has the greatest impact on system power consumption, but it does not take into account the system load caused by changing the CPU frequency and GPU frequency. balanced situation. By decomposing a parallel program into tasks that can be executed simultaneously, and mapping each task to the most suitable processor, the computing power of the system can be fully utilized to improve system performance, but this mapping scheme usually does not take system power consumption into consideration. Therefore, the present invention combines two optimization strategies of DVFS and task mapping to limit system power consumption to a certain budget level while improving system performance.

Description of drawings

In order to make the purpose of the present invention, the scheme more easy to understand, the present invention will be further described below in conjunction with accompanying drawing.

Figure 1 is a CPU-GPU heterogeneous multi-core system architecture diagram. The heterogeneous system is constructed by gem5-gpu simulation. A 4-core CPU and a GPU composed of 8 CUs are integrated on the same chip.

Fig. 2 is a schematic diagram of the design of a power capping scheme based on detailed operating information in the present invention.

Fig. 3 is a schematic diagram of fine-grained synchronization between CPU and GPU using traditional task data partitioning.

Fig. 4 is a schematic diagram of CPU and GPU busy time and CPU and GPU idle time caused by waiting for task management data of the task kernel function blocks used in the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

Figure 1 is a heterogeneous multi-core system built by the gem5-gpu simulator. It simulates a heterogeneous architecture that integrates a 4-core CPU and a GPU composed of 8 CUs on the same chip. In gem5-gpu , the simulation architecture can be flexibly modified according to the configuration file, and gem5-gpu supports DVFS.

In the heterogeneous multi-core system constructed by gem5-gpu simulation, the present invention realizes a power capping method by combining DVFS and task mapping, including the following specific steps:

Step 1, implement measuring the total computing node time, CPU power consumption and GPU power consumption after the execution of the application program is completed.

In gem5-gpu, after an application is executed, a file stat.txt containing all program execution information will be automatically generated, including the program execution time. The McPAT module can measure the CPU power consumption separately, and the GPUWattch module can measure the GPU module separately. By configuring the McPAT module and the GPUWattch module in gem5-gpu, the CPU power consumption and the GPU power consumption can be obtained after the program execution is completed.

Step 2, modify the selected parallel test benchmark program, and obtain the execution time of each kernel function in the program.

OpenCL programs can be executed on different devices, including CPUs and GPUs. The benchmark test program used in the present invention is the NAS parallel test benchmark program set of the OpenCL version. Each benchmark has different characteristics. Among them, some programs contain more than 60 kernels, and some programs only have two kernels. By rewriting the test program, the execution time of each kernel is collected.

Step 3: Execute the application program only on the CPU or the GPU respectively, set different CPU (GPU) optional frequencies, obtain detailed operation information, and use it as an input for a power capping scheme based on the operation information.

Figure 2 shows the flow of a power capping strategy based on operational information. Among them, CPU Profile Runs and GPUProfile Runs respectively represent the running information obtained after executing the application program only on the CPU and only on the GPU. Through these running information, the prediction model in step 4 is established, including the time model and the power consumption model. Represented by the Time model and Power model in Figure 2. The input of the prediction model is CPU frequency, GPU frequency and task mapping scheme, and the output is the corresponding predicted execution time and predicted system power consumption. Different input and output construct a configuration table, and use the improved greedy algorithm to search for the best mapping scheme and frequency setting according to the algorithm in step 6 according to the configuration table. Represented by Distribute parallel tasks and set device frequencies in Figure 2.

Step 4, establishing a prediction model, including an execution time model and a power consumption model, to predict the execution time and system power consumption of an application program in a heterogeneous multi-core environment. In step 3, under different frequency settings, the running information obtained by executing the application program only on the CPU and only on the GPU is the basis for establishing the prediction model. From formula ① to formula ⑧, it can be seen that through the program running information including the execution information of each kernel, CPU power consumption and GPU power consumption, the program execution time and power consumption.

The task mapping involved in the present invention refers to mapping any kernel in the program to the CPU or GPU as a whole, which is different from the traditional method of assigning a certain proportion of data to the CPU and GPU according to the task data. Allocation according to the proportion of task data means that a kernel executes the same code on the CPU and GPU at the same time, the data required by the kernel is allocated to the CPU and GPU in proportion, and the CPU and GPU process the allocated data at the same time. When assigning according to the task data ratio, since each kernel needs to be synchronized on the CPU and GPU after execution, the fixed data allocation ratio may cause a lot of idle CPU time and GPU time during kernel execution. As shown in Figure 3, the fine-grained synchronization between the CPU and GPU of the task data block is illustrated. The fixed data allocation ratio is α. For kernel1, the GPU execution efficiency is better, while the CPU processing speed will be slower. For the allocated data, the GPU will complete the execution before the CPU. At this time, the GPU is idle, waiting CPU execution completes. In contrast, for kernel2, the processing speed of the CPU is faster, so the CPU will complete the execution before the GPU. At this time, the CPU will be in an idle state and wait for the GPU to complete the execution. Therefore, it can be seen from Figure 3 that the CPU and GPU synchronization required for task data partitioning will cause a lot of idle time in the CPU and GPU during execution. For directly assigning a kernel as a whole to the CPU or GPU, there is no need for synchronization between the CPU and the GPU, but relatively, the data transfer time will be longer. As shown in Figure 4, it shows a CPU and GPU execution process based on kernel mapping as a whole, which does not require CPU and GPU synchronization, but data transmission between CPU and GPU is more frequent.

Step 5: Construct a configuration parameter table based on the prediction model.

Using the time prediction model and power consumption prediction model in step 4, the execution time and power consumption under all possible CPU frequencies, GPU frequencies and task mapping schemes can be calculated and stored in the configuration parameter table.

Step 6, according to the configuration parameter table, use the improved greedy algorithm to search for the optimal parameter set.

Step 6.1, given the CPU frequency and GPU frequency, search for the task mapping scheme with the shortest execution time, and obtain the predicted power consumption of the system according to the power consumption model. Shown by Algorithm 1.

Step 6.2, according to the optional settings of CPU frequency and GPU frequency, change the CPU frequency and GPU frequency, repeat step 6.1, obtain the mapping scheme corresponding to the optimal execution time under this frequency combination, and calculate the system power consumption according to this mapping scheme, Based on the given budget power consumption, the optimal frequency parameter selection and mapping scheme limited under the budget power consumption is obtained. This step is shown by Algorithm 2.

Claims (2)

1. A method for capping heterogeneous multi-core power by coordinating DVFS and task mapping, characterized in that, comprising the following steps: Step 1. Measure the total computing node time, CPU power consumption and GPU power consumption after the execution of the application program is completed; Step 2. Modify the selected parallel test benchmark program to obtain the execution time of each kernel function in the program; Step 3. Execute the application program only on the CPU or GPU respectively, set different CPU or GPU optional frequencies, and obtain detailed running information, including the total execution time of the program, the execution time of each kernel function in the program, and the total power consumption of the system , CPU power consumption and GPU power consumption; Step 4, design a predictive model to predict the power consumption and execution time of different task mapping schemes under different CPU and GPU frequencies; wherein, the input of the predictive model is CPU frequency, GPU frequency and task mapping scheme, and the task mapping refers to It is to map each kernel function as a whole to the CPU or GPU, instead of assigning the kernel function to the CPU and GPU according to a certain ratio for simultaneous execution; the output of the prediction model is the predicted total execution time of the program and the predicted system function. consumption; Step 5, calculate different CPU frequencies and GPU frequencies, execution time and power consumption under different task mapping schemes according to the execution time model and power consumption model, and fill in the configuration parameter table; Step 6, according to the configuration parameter table, use the improved greedy algorithm to search for the optimal parameter set; it is divided into two steps, first use the greedy algorithm to search for the task mapping scheme with the shortest execution time under the given CPU frequency and GPU frequency, according to this task Mapping, CPU frequency, GPU frequency, and power model to obtain the predicted power consumption of the system under this parameter configuration; then, change the CPU frequency and GPU frequency, and calculate the predicted power consumption of the system according to the previous step again, and finally get the power consumption limited to the budget The optimal configuration scheme below. 2. The heterogeneous multi-core power capping method by coordinating DVFS and task mapping as claimed in claim 1, wherein the prediction model in step 4 comprises: an execution time model and a power consumption model, Step 4.1, Execution Time Model The total execution time of the application program can be obtained according to the execution time of each core in the program and the corresponding data transmission time. The total execution time of the program is expressed by the formula ①. Among them, f _cpu and f _gpu represent the CPU frequency and GPU frequency respectively, and T _i (f _cpu , f _gpu ) represents the execution time of the i-th kernel function and the required data transmission time, expressed by formula ②. The first part of the formula represents the execution time, and the second part represents the data transfer time. The execution time of the kernel function has been obtained in step 2. H2D and D2H denote the data transfer cost from host to device and the transfer cost from device to host, respectively. Required is equal to 1 or 0, indicating whether the core k is required or the data d is required. The data may already be on the device and does not need to be transmitted, so use OnDevice to indicate whether the data is on the device. Use size to represent the data size. Step 4.2, Power Consumption Model System power consumption may be represented by three parts, which are idle power consumption, CPU power consumption P _cpu and GPU power consumption P _gpu . System power consumption is represented by formula ③. P＝P _idle (f _cpu ，f _gpu )+P _cpu (f _cpu ，f _gpu )+P _gpu (f _cpu ，f _gpu ) ③ Among them, P _idle represents idle power consumption, which is related to CPU frequency and GPU frequency, but has nothing to do with task mapping. It can be obtained by subtracting the power consumption on the CPU from the total power consumption obtained by executing the application program only on the CPU, and obtained by formula ④. Respectively represent the total power consumption of the system, CPU power consumption and GPU power consumption obtained under the condition that only the application program is executed on the CPU, and are related to the set frequency. When the application program is executed, the CPU and GPU are not always executing due to the data association between the kernel functions, so the CPU power consumption and GPU power consumption vary according to the given task mapping scheme, combined with this phenomenon , we assume that power consumption and execution time are proportional to the ratio of total execution time, so CPU and GPU power consumption can be expressed by formula ⑤ and formula ⑥ respectively. Among them, λ _cpu and λ _gpu represent the CPU and GPU busy time ratio, respectively, Respectively represent the CPU power consumption and the GPU power consumption obtained under the condition that the application program is only executed on the GPU. with estimators representing the maximum value of dynamic power consumption in the CPU and GPU, respectively. Since λ _cpu and λ _gpu respectively represent the ratio of CPU and GPU busy time, it can be obtained from the running information measured in step 3. When the program is executed, the busy time of each device is defined as the sum of the actual execution time of each kernel function, λ _cpu and λ _gpu can be represented by Equation ○7 and Equation 8○, respectively. t _cpu indicates the busy time of the CPU, t _gpu indicates the busy time of the GPU, and t _total indicates the total execution time of the application.