[zeromq-dev] EPGM & ZMQ_RATE

[Steven McCoy]

On 14 May 2011 00:19, Michael Holm <Michael.Holm at tibra.com> wrote:

> 1. Is anyone currently using EPGM/0MQ in a production environment that is
> willing to get in touch to discuss the performance.
>
> 2. I have been playing with the ZMQ_RATE option with EPGM. I have noticed
> that the higher the value the better the performance but when this value is
> set to 1,000,000 weird things happen with 0MQ on both sides. Typically a
> value of 100,000 seems to be okay and stable but the throughput is not as
> good.
>
> This is my test 100,000 messages (67 bytes) published to the clients.
>
> I see from the documentation the default value is 100 (kilobits per
> second). Is kilobits a typo?
>
>
That isn't a typo, but I did bump it up to 40mb/s (40 * 1000) in 2.1.6 to
make the default case more useful for quick testing, I wouldn't recommend
setting it above 600mb/s it would typically be safer to be at 40% line
capacity at 400mb/s.  The setting no longer includes recovery data as it
uses PGM_ODATA_MAX_RTE which is a more useful setting for most users, it can
be refined further with the OpenPGM API but it has not been shown necessary
to date.

http://code.google.com/p/openpgm/wiki/OpenPgm5CReferencePgm

You definitely want a rate set on Windows platforms as the platform really
does not like running at full speed, significantly driven by coarse grained
timing facilities.

I would quickly try some alternative timers to ensure your systems don't
have TSC instabilities, e.g.

PGM_TIMER=GETTIMEOFDAY

or

PGM_TIMER=CLOCK_GETTIME

http://code.google.com/p/openpgm/wiki/OpenPgmCReferenceRunWithTheseCapabilities

But it would look like your rate limit is way too high, consider performance
testing with pgmping and iperf to gain a base line.

-- 
Steve-o
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.zeromq.org/pipermail/zeromq-dev/attachments/20110514/1f7d3c26/attachment.htm>