[zeromq-dev] Async::Worker, C++ task offloading.

Martin Sustrik sustrik at 250bpm.com
Tue Jul 27 12:00:16 CEST 2010

Steven McCoy wrote:
> On 27 July 2010 16:59, Pieter Hintjens <ph at imatix.com 
> <mailto:ph at imatix.com>> wrote:
>     On Tue, Jul 27, 2010 at 10:44 AM, Steven McCoy <steven.mccoy at miru.hk
>     <mailto:steven.mccoy at miru.hk>> wrote:
>      > On 26 July 2010 18:50, Martin Sustrik <sustrik at moloch.sk
>     <mailto:sustrik at moloch.sk>> wrote:
>      >>
>      >> Given the complexity of benchmarking and high-perf optimisation,
>     letting
>      >> people post random benchmarks and tips would just cause
>     confusion IMO.
>      >>
>      >
>      > Definitely as ICC is actually slower for PGM.
>     Isn't that the kind of knowledge that's worth collecting?  Why do you
>     say that letting people post random benchmarks and tips would just
>     cause confusion?  We do this all the time on the mailing list, and I
>     don't see any confusion.
> As long as there is one reference platform used for every release 
> somewhere that can provide an indication of performance from one release 
> to the next it would be fine.  The problem is that messaging performance 
> is always dependent upon so many factors that individual benchmark 
> results have little meaning.
> You could probably run through the last dozen kernel releases on the 
> same platform and get very different result sets.
> I don't like TIBCO's stance of never publishing any results, or Reuters 
> stance of publishing completely useless wire speed results as they have 
> no practical value.
> If you can use a standard boring Dell and CentOS and run a heat map 
> graph of results for 4/5/6 and each set of patterns it would be great. 
>  I wouldn't want to limit contributed results for alternative platforms 
> such as QNX, ARM, etc, but there is a value to one core set of readings.

Fully agreed with the above.


More information about the zeromq-dev mailing list