pthreads don't scale on windows xp, but does scale on linux, cygwin 1.5.19

[jdeifik]

At 07:00 PM 3/21/2006, you wrote:

>jdeifik wrote:
>>I have a dual xeon 2.4ghz machine with hypertreading enabled.
>>This gives me 4 logical processors.
>>The machine dual boots to windows xp sp2, and linux.
>>I have a highly parallelizable program I wrote, and I tested it 
>>running 1 to 8 threads,
>>running with no source changes on windows and linux.
>>Here is the performance on linux using gcc-3.4.3
>>threads
>>1    1436.41user 0.10system 7:16.37elapsed 100%CPU 
>>(0avgtext+0avgdata 0maxresident)k
>>2    436.00user 0.02system 3:38.15elapsed 199%CPU 
>>(0avgtext+0avgdata 0maxresident)k
>>3    369.15user 0.05system 2:03.48elapsed 298%CPU 
>>(0avgtext+0avgdata 0maxresident)k
>>4    359.77user 0.08system 1:42.95elapsed 349%CPU 
>>(0avgtext+0avgdata 0maxresident)k
>>6    357.83user 0.09system 1:40.94elapsed 354%CPU 
>>(0avgtext+0avgdata 0maxresident)k
>>8    358.79user 0.06system 1:41.80elapsed 352%CPU 
>>(0avgtext+0avgdata 0maxresident)k
>>To compute efficiency, take the single thread elapsed time/(# 
>>threads * threaded elapsed time)
>>There is virtually perfect scaling. 4 processors scale with an 
>>efficiency of about 103%.
>>For 6 and 8 threads, efficiency goes up a small amount.
>>
>>Here is the performance on windows xp using cygwin pthreads and gcc-3.4.4
>>1    434.60user 0.20system 7:16.47elapsed 99%CPU (0avgtext+0avgdata 
>>509696maxresident)k
>>2    441.78user 0.24system 3:42.06elapsed 199%CPU 
>>(0avgtext+0avgdata 510208maxresident)k
>>3    579.68user 0.15system 3:14.50elapsed 298%CPU 
>>(0avgtext+0avgdata 511232maxresident)k
>>4    675.39user 0.15system 2:51.50elapsed 393%CPU 
>>(0avgtext+0avgdata 512000maxresident)k
>>6    711.70user 0.18system 3:01.20elapsed 392%CPU 
>>(0avgtext+0avgdata 511488maxresident)k
>>8    683.35user 0.21system 2:56.05elapsed 388%CPU 
>>(0avgtext+0avgdata 512000maxresident)k
>>Things are fine for 2 threads, scaling with an efficiency of 96%
>>For 3 threads, scaling efficiency is 73%
>>For 4 threads, scaling efficiency is 62%
>>For 6 threads, scaling efficiency is 39%
>>For 8 threads, scaling efficiency is 30%
>
>Windows doesn't have HT aware scheduling, such as recent linux 
>schedulers incorporate.  Cygwin doesn't attempt to improve on the 
>Windows scheduler.  I won't ask for relevant details about your 
>linux, or how you managed to write a program which doesn't deliver 
>close to full performance at 2 threads, as that would take this even 
>further Off Topic. However, if you are getting good scaling to 2 
>threads, that should enable you to get all the dual processor 
>performance you can expect in Windows for practical purposes.  You 
>might try repeating your tests with HT disabled in BIOS.

My linux is mandrake 10.2, I suspect running kernel 2.6.11-13smp.
My program scales perfectly at 2 threads on linux. It also scales 
perfectly at 4 threads on linux.
The problem isn't with my program.

I am not sure why it is important to have a HT aware scheduler for 
Windows, when there are 4 or more
threads. I can see with 2 threads you would like to have one per 
physical processor.
With 4 or more threads, cygwin phtreads really sucks, 4->62%, 6->39%, 
8->30% efficiency.

It seems to me that more and more apps are turning to threading for 
performance,
and more and more hardware is available with multi-processor, 
multi-core, and multi-threading.

         Jeff Deifik 


--
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple
Problem reports:       http://cygwin.com/problems.html
Documentation:         http://cygwin.com/docs.html
FAQ:                   http://cygwin.com/faq/