Le 06/08/2014 08:10, jhu a écrit :
> clipka <ano### [at] anonymousorg> wrote:
>> Am 06.08.2014 03:41, schrieb jhu:
>>> "green" <rov### [at] gmailcom> wrote:
>>>> something i found today
>>>> http://www.pugetsystems.com/blog/2014/07/14/POV-ray-on-Quad-Xeon-and-Opteron-579/
>>>> -or- http://tinyurl.com/mkc3bkz
>>>
>>> That is interesting. Of course, the Linux version is faster because it's a
>>> custom compile. What I'm surprised about is that hyperthreading kills
>>> performance on Windows for some reason. Can anyone explain why?
>>
>> As far as I understand it doesn't really kill performance, it just
>> doesn't add much.
> 
> Look at the bottom. It definitely kills performance on Windows.
> 
> 
The windows scheduler sucks to detect HT-core from true core.

The linux scheduler is "smarter", when you have 40 true core and 40
HT-core, if you have to run 30 threads, it will choose distinct true+HT
core pairs for each threads.

I guess, if we set HT-core as odd and true core as even, that the
windows scheduler when he has 30 threads over 40+40 cores would use the
0-29 range, actually using 15 true cores and 15 associated HT-core, thus
leaving 25 true cores idles (and their peer HT-core too). With a bit of
help, other processes in the system might shift/expand the range of used
cores. Yet, with HT active, the distribution is wasting resources and
creating bottleneck.

Now the blogging guy is just lucky to have such beasts (yep, he has 2!)

-- 
Just because nobody complains does not mean all parachutes are perfect.

Post a reply to this message