UltraSPARC T2 SPECcpu autopar
By sprack on Nov 13, 2008
Following from previous discussions on the benefits of autopar for SPECcpu, the next logical question is "what kind of benefit does autopar provide on CMT processors like the UltraSPARC T2". On UltraSPARC T2, we have a multitude of hardware strands available and, as previously discussed, the bare-metal inter-thread communication latencies are extremely low. I talked with some of the compiler gurus at Sun and sure enough this analysis had been undertaken for SPECfp. The results are as follows:
Pretty cool! 7 of the suites X benchmarks show some benefit, with 1 showing over 30X speedup, a further 3 showing a benefit of over 10X, and the remainder showing 2-4X improvements. Some of the benchmarks show peak performance when using less than 64-threads. This is not unexpected, as this is an out-of-the-box run and given that T2 has shared pipelines and, like most multi-core processors, shared caches and offchip bandwidth, some tweaking is required to maximize performance.