Experimental analysis of operating system jitter caused by page reclaim

Operating system jitter is one of the causes of runtime overhead in high-performance computing applications. Many high-performance computing applications perform burst accesses to I/O, and such accesses consume a large amount of memory. When the Linux kernel runs out of memory, it awakens special kernel threads to reclaim memory pages. If the kernel threads are frequently awakened, application performance is degraded because of the threads’ resource consumption as well as the increase in the application’s page faults and migration between CPU cores. In this study, we empirically analyze the impact of jitter caused by reclaiming memory pages, and we propose a method for reducing it. The proposed method reclaims memory pages in advance of the kernel thread. It reclaims more pages at one time than the kernel threads, thus reducing the frequency of page reclaim and the impact of jitter. We conducted experiments using practical weather forecast software, the results of which showed that the proposed method minimized performance degradation caused by jitter.