qpid-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Carl Trieloff <cctriel...@redhat.com>
Subject Re: AMQP Paper, using Qpid
Date Wed, 12 Nov 2008 02:13:03 GMT
Carl Trieloff wrote:
>
> Greg,
>
> I was pointed to a link to this paper...,  a nice piece of work:
>
> http://www.cse.ohio-state.edu/~narravul/papers/subramoni_whpcf08.pdf
>
> Few comments that may help you if you are still doing work in the area.
>
> - You should be able to achieve better results by using the raw OFED 
> RDMA support Qpid has
> - There has been quite a lot of mutex work in the broker, which I 
> believe is post the report.
> - there have been updates around consumer side threading, which I 
> believe is also post the report.
> - On the IB side, RHEL 5.2 is OFED 1.3, that should also give you 
> better results.
> - You talk about decentralization, you could look at these cases using 
> Federation, note that federation
> also supports OFEB 1.3 RDMA in Qpid.
>
> As the report is in the context of financial applications, there are a 
> few additional patterns that
> I would be glad to provide information on. These are the support for 
> LVQ, Ring queue, and LVE, etc
>
> Few area's that have been played with that would maybe also be of 
> interest to your research are:
> - lock-free queues impl for the broker. Have seen some benefits in 
> this area
> - there are some interesting changes coming down from upstream Linux 
> around private futextes which may
> have significat effect, there is also some changes to Nagle options 
> happening in upstream Linux. you would
> need to pick up RHEL 5.3 to play with the futexes, or MRG-RT.
>
> Nice piece of work, happy to provide any additional information for 
> your future work section
> regards,
> Carl.


"Our analysis leads us to believe that the IPoIB limitations
seen in our tests are due to the high IPoIB stack overhead.
We believe that applying modern communication protocols,      
like RDMA, would improve this performance. " 

from the report ---  I can confirm this, as the IPoIB driver consumers
considerable CPU, which RDMA OFED does not.

also, if you want to measure latency, setting
--max-frame-size
--tcp-nodely
--bounds-multiplier

on the client will greatly affect latency
Carl.




Mime
View raw message