Hyperthreading degrades performance?
andre at freebsd.org
Thu Aug 25 18:12:12 GMT 2005
Robert Watson wrote:
> On Thu, 25 Aug 2005, Justin R. Smith wrote:
>> This is in reply to the people who said this because of
>> cache-contention. Has anyone benchmarked this?
>> There's an article
>> that benchmarks hyperthreading in Linux and shows a modest (~29%)
>> improvement in performance --- depending on applications (with java
>> showing degradation of performance). Perhaps the linux sheduler does
>> things differently...
> In in the last couple of years, I've seen some changes in how we
> interact with HTT. 2-3 years ago, when benchmarking MySQL with and
> without HTT, I saw a 30%+ drop-off when HTT was enabled. Now, they come
> out about the same. I previously also saw no improvement with
> buildkernel, but recently I've seen credible reports of build
> improvements when running with HTT. So I think that things have changed
> a bit as a result of significant scheduler improvements in the last few
> years, as well as reduced lock contention. A continued slight decrease
> in performance for some benchmarks wouldn't surprise me, but seeing more
> in the way of "break even" or even "improvement" strikes me as likely.
> A thorough revisiting of the issue would be quite useful :-).
Don't forget better PIV revisions with larger instruction decoder caches,
better cache prefetching and branch prediction. I doubt much of the
improvement is due to our SMP changes. A real test to find out whether
it's our work or Intels would be to benchmark an old (pre Nacona) PIV
running 5.3R and 7.0-current vs. a new one doing the same.
More information about the freebsd-current