lkml.org 
[lkml]   [2006]   [Jan]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch 2/3] mutex subsystem: fastpath inlining

Sorry for the delay...

On Thu, 29 Dec 2005, Ingo Molnar wrote:

>
> * Nicolas Pitre <nico@cam.org> wrote:
>
> > This is with all mutex patches applied and CONFIG_DEBUG_MUTEX_FULL=n,
> > therefore using the current semaphore code:
> >
> > text data bss dec hex filename
> > 1821108 287792 88264 2197164 2186ac vmlinux
> >
> > Now with CONFIG_DEBUG_MUTEX_FULL=y to substitute semaphores with
> > mutexes:
> >
> > text data bss dec hex filename
> > 1797108 287568 88172 2172848 2127b0 vmlinux
> >
> > Finally with CONFIG_DEBUG_MUTEX_FULL=y and fast paths inlined:
> >
> > text data bss dec hex filename
> > 1807824 287136 88172 2183132 214fdc vmlinux
> >
> > This last case is not the smallest, but it is the fastest.
>
> i.e. 1.3% text savings from going to mutexes, and inlining them again
> gives up 0.5% of that. We've uninlined stuff for a smaller gain in the
> past ...
>
> > > Note that x86 went to a non-inlined fastpath _despite_
> > > having a compact CISC semaphore fastpath.
> >
> > The function call overhead on x86 is less significant than the ARM
> > one, so always calling out of line code might be sensible in that
> > case.
>
> i'm highly doubtful we should do that. The spinlock APIs are 4 times
> more frequent than mutexes are ever going to be, still they too are
> mostly out of line. (and we only inline the unlock portions that are a
> space win!) Can you measure any significant difference in performance?
> (e.g. lat_pipe triggers the mutex fastpath, in DEBUG_MUTEX_FULL=y mode)

Here it is.

With the default non inlined fast path:

Pipe latency: 14.2669 microseconds
Pipe latency: 14.2625 microseconds
Pipe latency: 14.2655 microseconds
Pipe latency: 14.2670 microseconds

Then, with fast paths inlined:

Pipe latency: 13.9483 microseconds
Pipe latency: 13.9409 microseconds
Pipe latency: 13.9468 microseconds
Pipe latency: 13.9529 microseconds

So inlining the mutex fast path is more than 2% faster for the whole
test.

Is that worth the 0.5% increase in kernel size in your opinion?

Note: I modified lat_pipe to use two threads instead of two processes
since on ARM switching mm require a full cache flush due to the VIVT
cache, and doing so skyrockets the pipe latency up to 480 microsecs with
the mutex difference lost in the noise.


Nicolas
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2006-01-06 22:23    [W:0.090 / U:0.112 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site