Analyzing the characteristics of memory subsystem on two different 8-way NUMA architectures

Abstract

Two NUMA architectures with different memory subsystems are experimentally analyzed in this paper. By applying the benchmark with various access patterns, it shows much different characteristics of memory system between Xeon E5620 with Global Queue and LS 3A with typical crossbar switch. The experiment results reveal the fact that LS 3A and Xeon E5620 have some similar features. Our study also showed some other diverse features of these two platforms: due to the different contention locations and mechanisms, the memory access model for E5620 doesn't fit for LS 3A. Through comparing, we find that one advantage of LS 3A is that it can obtain steady bandwidth on both local and remote thread, and it is more fair for local and remote access under some circumstances. Another fact is that LS 3A is not such sensitive to remote access, compared with E5620, so there will be no obvious performance degradation caused by non-local memory access. © 2013 IFIP International Federation for Information Processing.