qpid-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Gordon Sim <g...@redhat.com>
Subject Re: multiple listener ahsring a public queue
Date Thu, 20 Nov 2008 16:43:43 GMT
mark yoffe wrote:
> Hi
> i have been running scenario where several consumers connect to the same
> public queue (topic)
> the consumers connect via subscription manager.
> i have run several scenarios where multiple listeners use the same
> connection (different sessions) and scenario where listeners each have their
> own connection
> all of this has been run on TRUNC versions
> a while ago i started experiencing a serious down downgrade in performance
> when using such a scenario
> in the past the use of multiple listeners enabled me to enhance performance
> but currently it appears that i can achieve a faster performance while using
> 1 listener application
> in the past when several listeners used a shared queue - the message flow
> to between the consumers was not even  (if using x listener one would expect
> to receive 1/x messages per listener)
> this problem/feature was corrected to support a pretty even distribution of
> the task
> as far as i can tell around the time this enhancement was made(it might not
> be connected but..)
> the performance started going down, and it is not a small difference
> using one listener i can achieve a speed of X for 100k messages per second
> using five listeners i experience a serious down grade of performance to ~5X
> for 100k messages per second
> this was not the case in the past and i am at a loss regarding this problem
> does anyone know , how this can be resolved

Sounds like https://issues.apache.org/jira/browse/QPID-1280 for which I 
don't have a quick fix unfortunately as yet. It was caused by some 
necessary changes to the locking (specifically to prevent deadlocks when 
using RDMA, but this may also have improved the 'fairness' of allocation).

I can give you a one-line patch that should revert back to the earlier 
performance if you like. I have some ideas for a proper fix but they 
were going to be too involved for the M4 timeframe I'm afraid.

Adding more publishers also helps I found.

View raw message