An implementation of fast memset() using hardware accelerators: extended abstract

Abstract

Multicore systems with large caches and huge main memories have become ubiquitous. They provide an attractive opportunity to maximize performance of big-memory applications such as in-memory databases, key-value stores, and graph analytics. However, these big-memory applications require many virtual-to-physical address translations, which increase TLB miss rate and hurt performance. To address this problem, modern hardware and OSes introduced support for huge pages. For example, on SPARC M7, Linux supports 8MB, 2GB, and 16GB huge pages (in addition to the default 8KB). Likewise, Linux supports 2MB and 1GB huge pages on Intel Xeon (E5-2630) platforms.