lkml.org 
[lkml]   [2009]   [Jan]   [14]   [last100]   RSS Feed
Views: [more markup]  [less markup]  [headers]  [forward] 
 
Messages in this thread
/
DateWed, 14 Jan 2009 20:28:11 +0100
FromIngo Molnar <>
SubjectRe: [GIT PULL] adaptive spinning mutexes
* Ingo Molnar <mingo@elte.hu> wrote:

> Latest performance figures, on a 2-socket 16-way Nehalem test-system, 
> running the code above, measured via "test-mutex V 128 10" VFS 
> creat+unlink scalability test on tmpfs and ext3:
> 
>                               no-spin           spin
> 
>  [tmpfs]    avg ops/sec:       291038         392865       (+34.9%)
>  [ext3]     avg ops/sec:       283291         435674       (+53.7%)

Btw., for historic kicks i just went back to v2.6.15-2019-gf17578d - the 
last pre-mutexes semaphore based kernel, using the same .config.

I tracked down two bugs in it to make it boot on a Nehalem, so we can now 
compare the above numbers against historic semaphore performance:

                      [v2.6.14]     [v2.6.29]

                      Semaphores  | Mutexes
            ----------------------------------------------
                                  | no-spin           spin
                                  |
  [tmpfs]   ops/sec:       50713  |  291038         392865       (+34.9%)
  [ext3]    ops/sec:       45214  |  283291         435674       (+53.7%)
A 10x macro-performance improvement on ext3, compared to 2.6.14 :-)

While lots of other details got changed meanwhile, i'm sure most of the 
performance win on this particular VFS workload comes from mutexes.

So i think the long mutex migration pain was definitely worth it.

	Ingo


\
 
 \ /
  Last update: 2009-01-14 20:31    [from the cache]
©2003-2008