RE: Performance: C++ client - Windows VS LInux

[Bruno Matos <br...@paradigmaxis.pt>]

Hi Steve,

On Fri, 2010-09-24 at 11:48 -0400, Steve Huston wrote:
> Hi Bruno,
> 
> > Hi Steve,
> > 
> > On Fri, 2010-09-24 at 10:49 -0400, Steve Huston wrote:
> > > These look like profiling info from the client side, is 
> > that right? It 
> > > shows a lot of waiting, which indicates the delays are 
> > probably in the 
> > > broker side. Could you please profile the broker while running your 
> > > timing test?
> > 
> > Yes, I can do the profiling in broker, but notice that I use 
> > the same broker all the time. Do you think it still helps? 
> > The broker is running on Fedora 13.
> 
> Ok, I forgot about this... No, don't bother profiling the broker. But it
> would be useful to profile the client again, but run a few thousand
> messages to make sure that the message-processing path stands out from
> the noise.

In attachment is a sleepy profile of 1.000.000 messages.

> 
> Thanks,
> -Steve
> 
> > > > -----Original Message-----
> > > > From: Bruno Matos [mailto:bruno.matos@paradigmaxis.pt]
> > > > Sent: Friday, September 24, 2010 9:36 AM
> > > > To: dev@qpid.apache.org
> > > > Subject: RE: Performance: C++ client - Windows VS LInux
> > > > 
> > > > 
> > > > On Fri, 2010-09-24 at 11:49 +0100, Bruno Matos wrote:
> > > > > On Thu, 2010-09-16 at 07:50 -0400, Steve Huston wrote:
> > > > > > Hi Bruno,
> > > > > > 
> > > > > > > Thanks for the reply.
> > > > > > 
> > > > > > You're welcome.
> > > > > > 
> > > > > > > On Mon, 2010-09-13 at 07:06 -0400, Steve Huston wrote:
> > > > > > > > Hi Bruno,
> > > > > > > > 
> > > > > > > > > I'm facing some performance issues with a 
> > Windows client. 
> > > > > > > > > I
> > > > > > > > > made some tests and the difference is between 625857 
> > > > > > > > > microseconds/packet in Windows and 30110 
> > > > microseconds/packet
> > > > > > > > > in Linux. This is the average of 10.000.000 packets.
> > > > > > > > > 
> > > > > > > > > The windows libs were compiled from svn tag 0.6,
> > > > and the Linux
> > > > > > > > > libs are from Fodera 13's Yum repos.
> > > > > > > > 
> > > > > > > > What version of Qpid did you test with on Fedora?
> > > > > > > 
> > > > > > > I'm using 0.6.
> > > > > > > 
> > > > > > > > 
> > > > > > > > > Some ideas? There are any precopiled distribution
> > > > supported by
> > > > > > > > > the project?
> > > > > > > > 
> > > > > > > > You can get a Qpid 0.6 installable from 
> > > > > > > > http://www.riverace.com/qpid/downloads.htm, but it's
> > > > 0.6 - probably
> > > > > > > > not significantly different from what you tested.
> > > > > > > > 
> > > > > > > 
> > > > > > > With this libs I get 573361 microseconds/packet. A little 
> > > > > > > less, but not enough.
> > > > > > 
> > > > > > Right...
> > > > > > 
> > > > > > > > If you get profiling info that may help to improve this,
> > > > > > > please let me
> > > > > > > > know. I'm also available to help get that information.
> > > > 
> > > > I have two output files from Sleepy, with asynchronous and 
> > > > synchronous session. Sleepy can be found in 
> > > > www.codersnotes.com/sleepy.
> > > > 
> > > > > > > 
> > > > > > > I think I will do a simple test program only for this. What
> > > > > > > do I need to get useful profiling information?
> > > > > > 
> > > > > > Output from any reasonable performance measurement. Something 
> > > > > > like Rational Quantify would do it, or one of Intel's thread
> > > > measuring tools
> > > > > > (I forget the name).
> > > > > > 
> > > > > 
> > > > > I have a simple program now that sends and receives
> > > > messages. It starts
> > > > > counting time right before sending (synchronous), and stop
> > > > counting when
> > > > > message arrives. I get 757 microseconds/packet in Linux 
> > and 39118 
> > > > > microseconds/packet in Windows as an average of 10.000
> > > > packets. I will
> > > > > try Very Sleepy free profiling tool.
> > > > 

Regards,
Bruno

RE: Performance: C++ client - Windows VS LInux

[Bruno Matos <br...@paradigmaxis.pt>]

Hi,

Resume of previous conversations:
There is a big difference handling messages from a Linux C++ Qpid
broker, between Windows and Linux. Windows client is about 20x slower.

I made some profiling and Steve Huston is helping me detecting the
problem.

Can anyone help too?

Thank you,
Bruno Matos




---------------------------------------------------------------------
Apache Qpid - AMQP Messaging Implementation
Project:      http://qpid.apache.org
Use/Interact: mailto:dev-subscribe@qpid.apache.org