[Ace-users] Re: [tao-users] performance of COS Notification

[Friedhelm Wolf]

On 7/24/07, Matthew Gillen <mgillen at bbn.com> wrote:
>
> Friedhelm Wolf wrote:
> > 6. I did some measurements and found that even if the latency for 15
> > consumers is different for every consumer (basically it looks like that
> > the calls are processed serially and the last consumer has to wait the
> > longest to get an answer) the throughput remains the same. I'm not sure
> > if I understand it the right way, but should a bad latency not result in
> > reduced throughput?
>
> Not necessarily.  Latency and throughput are often related, but not
> always.
> In this case, it looks like some consumers were always 'first in line' to
> get events, and thus had lower latencies (supposing latency is measured
> from
> publisher-event-generation-time to consumer-event-delivery-time).


Ok, so that means, that the events in the tests are somehow sent serialized,
right?
And the reason, why this is not affecting throughput (at least not
significantly) is,
that the latency is relativly small (mean below 2500 us) compared to the
cycle
time of sending events, which is 10 ms, so there is enough time to send all
packets
before the next burst?

What's kind of strange about your results is that the publisher had a
> slightly lower publish throughput (410.17 event/sec) than all the
> consumer's
> throughput (410.22 event/sec):


Which means, that the publisher is sends less data, than is received by the
consumers?

Thanks for your comments. As I told, it is not clear to me, how to interpret
the
numbers, so your hints are very welcome.

Cheers,
Friedhelm
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://list.isis.vanderbilt.edu/pipermail/ace-users/attachments/20070725/01058dc1/attachment.htm