Evaluating the Effect of Huge Page on Large Scale Applications

Abstract

The ever-increasing memory footprint in large scale applications not only requires efficient computation, but also needs efficient communication. Traditionally huge page is used to reduce CPU's TLB miss, but how huge page impacts communication performance, and how HPC applications in large scale benefit from huge page are still not clear. In this paper, we support huge page in our D5000A platform and micro-benchmark the communication impacts of huge page in detail, also we evaluate the huge page impact on large scale HPC applications. Results show that huge page improves communication bandwidth by 40MB/s in large message sizes, greatly benefits communication performance when multiple connections share the same NIC. Applications such as NPB, Sweep3D, HPL show good performance improvement and good scalability in large scale. With huge page support, HPL achieves 2.76% improvement in 16384 cores.