[zeromq-dev] Sun Studio on Linux assembler

Steven McCoy steven.mccoy at miru.hk
Fri May 28 03:54:59 CEST 2010


On 28 May 2010 07:12, Martin Lucina <mato at kotelna.sk> wrote:

> I have a feeling that when I last looked Intel C did support proper GCC
> inline asm. icc is actually damn good, and can get nice performance
> improvements for some scenarios.
>
>
Completely invisible when using "-gcc-version=420".  For a network protocol
like PGM it makes absolutely zero improvement as everything is IO bound.
 Interesting any optimisations aside of "-O2" make it worse, "-mtune=native"
appears transparent but "-march=native" drops performance on x86-64.

Changing the memory allocator also is usually bad, for instance Google's
tcmalloc is worse.  I've been performance testing the difference between
zero-copy and copy in OpenPGM and it doesn't appear to affect performance
which would imply the zero-copy simply frees up FSB time for the
application.

If OpenPGM can copy SKBs then it can use a chunked ring buffer allocator
which improves performance by removing the one alloc per packet overhead.
 Ideally though the libc allocator thread caching should be sufficient to
reduce the benefit of a custom allocator.

-- 
Steve-o
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.zeromq.org/pipermail/zeromq-dev/attachments/20100528/20ad6451/attachment.htm>


More information about the zeromq-dev mailing list