SPECcpu and autopar
Browsing a few internet forums, it interesting to note that there is quite a debate on the use of autopar in SPEC results, with some calling for its use to be disallowed. While it is true that when autopar is permitted, the resulting performance is no longer single-thread performance, it nevertheless it an interesting measure of how well a particular processor/compiler combination perform (Even before the common use of autopar, SPEC was never just a measure of performance, but was heavily influenced by the sophistication of the compiler). Further, in today's multi-core processors, where caches & off-chip bandwidth are shared, even single-thread SPEC runs don't give all the of information one needs to fully understand the impact on the processor. Rather, SPEC ratio is merely an indication of of well the processor/system can run a single-instance of an application (i.e. peak single-application performance) and SPECrate is a measure of the peak throughput.
Finally, it can be argued that many HPC codes are amenable to autopar, so its use in SPECfp is relevant. Its use in SPECint is more problematic as most integer codes are more difficult to noticeably accelerate using autopar -- its just unfortunate that SPEC06int includes libquantum.....
