Hmem: A Holistic Memory Performance Metric for Cloud Computing

Yuyang Li, Ning Li, Yilei Zhang, Jianmei Guo, Bo Huang, Mengbang Xing, Wenxin Huang

Abstract: With the proliferation of cloud computing, cloud service providers offer users a variety of choices in terms of pricing and computing performance. A critical factor impacting computing performance is main memory, often evaluated using bandwidth and access latency metrics. For two evaluations with the same workload while under different system configurations, it is hard to determine which system delivers better memory performance for the particular workload if neither evaluation data achieves higher bandwidth and lower latency simultaneously. This dilemma is further exacerbated under different memory access patterns. We recognize that state-of-the-art memory performance metrics cannot well address the dilemma. To address this challenge, we define a holistic memory performance metric, named Hmem, which is calculated from a fusion of bandwidth and latency metrics across different access patterns. To reflect the overall performance of a given workload, we calculate the correlation between our proposed metric and the workload’s throughput. Experimental results show that Hmem exhibits an average improvement of 70% on correlation coefficients compared to state-of-the-art memory performance metrics. A large cloud service provider has adopted Hmem to improve the efficiency of their memory performance evaluation and cloud server selection.