To hardware prefetch or not to prefetch?

Abstract

Most hardware and software venders suggest disabling hardware prefetching in virtualized environments. They claim that prefetching is detrimental to application performance due to inaccurate prediction caused by workload diversity and VM interference on shared cache. However, no comprehensive or quantitative measurements to support this belief have been performed. This paper is the first to systematically measure the influence of hardware prefetching in virtualized environments. We examine a wide variety of benchmarks on three types of chip-multiprocessors (CMPs) to analyze the hardware prefetching performance. We conduct extensive experiments by taking into account a number of important virtualization factors. We find that hardware prefetching has minimal destructive influence under most configurations. Only with certain application combinations does prefetching influence the overall performance. To leverage these findings and make hardware prefetching effective across a diversity of virtualized environments, we propose a dynamic prefetching-aware VCPU-core binding approach (PAVCB), which includes two phases - classifying and binding. The workload of each VM is classified into different cache sharing constraint categories based upon its cache access characteristics, considering both prefetch requests and demand requests. Then following heuristic rules, the VCPUs of each VM are scheduled onto appropriate cores subject to cache sharing constraints. We show that the proposed approach can improve performance by 12% on average over the default scheduler and 46% over manual system administrator bindings across different workload combinations in the presence of hardware prefetching.