SPECrate2006: Alternatives Considered, Lessons Learned

* Final gross prices may vary according to local VAT.

Get Access

Abstract

Since 1992, SPEC has used multiple identical benchmarks to measure multi-processor performance. This “Homogeneous Capacity Method” (aka “SPECrate”) has been criticized on the grounds that real workloads are not homogeneous. Nevertheless, SPECrate provides a useful window into how systems perform when stressed by multiple requests for similar resources. This paper reviews SPECrate’s history, and several performance lessons learned using it: (1) a 4:1 performance gain for startup of a benchmark when I/O was reconfigured; (2) a benchmark that improved up to 2:1 when a TLB data structure was re-sized; and (3) a benchmark that improved by 52% after a change to NUMA page allocation. The SPEC CPU workloads usefully exposed several opportunities for performance improvement.