Optimal GPU Frequency Selection using Multi-Objective Approaches for HPC Systems
Optimal GPU Frequency Selection using Multi-Objective Approaches for HPC Systems
复制标题
DOI：
10.1109/hpec55821.2022.9926317
复制
发表时间：
2022-09
期刊：
影响因子：
--
通讯作者：
Ghazanfar Ali;Sridutt Bhalachandra;N. Wright;Mert Side;Yong Chen
中科院分区：
文献类型：
--
作者：
Ghazanfar Ali;Sridutt Bhalachandra;N. Wright;Mert Side;Yong Chen
功耗对当前和新兴的支持GPU的高性能计算（HPC）系统构成了重大挑战。在现代GPU中，存在诸如动态电压频率缩放（DVFS）等控制来调节功耗。由于不同的计算强度和广泛的频率设置的可用性，为给定的GPU工作负载选择最佳频率配置是不平凡的。应用具有降低功率的单一目标的功率控制可能导致性能降级，从而导致更多的能量消耗。在这项研究中，我们表征和识别GPU利用率指标，影响给定工作负载的功率和执行时间。分析模型的功率和执行时间，然后提出使用的特征集。多目标函数（即，能量延迟乘积（EDP）和ED2p）用于为工作负载选择最佳GPU DVFS配置，使得在没有性能降级或性能降级可忽略的情况下降低功耗。该评估是在NVIDIA GV100 GPU上使用SPEC ACCEL基准进行的。所提出的功率和性能分析模型的预测精度分别高达99.2%和98.8%。平均而言，基准测试显示，使用EDP和ED2p方法分别节省了28.6%和25.2%的能源，而没有性能下降。此外，所提出的模型只需要在最大频率，而不是所有支持的DVFS配置的度量收集。
Power consumption poses a significant challenge in current and emerging GPU-enabled high-performance computing (HPC) systems. In modern GPUs, controls like dynamic voltage frequency scaling (DVFS), among others, exist to regulate power consumption. Due to varying computational intensities and the availability of a wide range of frequency settings, selecting the optimal frequency configuration for a given GPU workload is non-trivial. Applying a power control with the single objective of reducing power may cause performance degradation, leading to more energy consumption. In this study, we characterize and identify GPU utilization metrics that influence both the power and execution time of a given workload. Analytical models for power and execution time are then proposed using the charac-terized feature set. Multi-objective functions (i.e., energy-delay product (EDP) and ED2p) are used to select an optimal GPU DVFS configuration for a workload such that power consumption is reduced with no or negligible degradation in performance. The evaluation was conducted using SPEC ACCEL benchmarks on NVIDIA GV100 GPU. The proposed power and performance analytical models demonstrated prediction accuracies of up to 99.2% and 98.8%, respectively. On average, the benchmarks showed 28.6% and 25.2% energy savings using EDP and ED2p approaches, respectively, without performance degradation. Fur-thermore, the proposed models require metric collection at only the maximum frequency rather than all supported DVFS configurations.