[zeromq-dev] epgm performance numbers

[Rohan Bedarkar]

Steve

Throttling definitely helps. I reduced the HWM to 200 and added a 200 ms sleep every 50 samples and that brought down average round-trip times massively...

So why is throttling helping? 

---------------------------------------------------------
             Performance Report                          
---------------------------------------------------------
                   
 Transport:                     ZMQ
 Protocol:                      epgm
 Messages sent:                1000
 Average / message:       175.561 us   
 Max time:                      5054 us   
 Min time:                      143 us   




 
--

Rohan Bedarkar
MS, Computer Science 

University of Chicago
rohanb at cs.uchicago.edu


 




________________________________
 From: Rohan Bedarkar <r_bedarkar at yahoo.com>
To: ZeroMQ development list <zeromq-dev at lists.zeromq.org> 
Sent: Monday, October 29, 2012 2:19 PM
Subject: Re: [zeromq-dev] epgm performance numbers
 

Steve,

I found something interesting. ZMQ roundtrip over epgm drops suddenly at one point.. Thoughts?

........
......
...
88: 143us
89: 144us
90: 145us
91: 146us
92: 147us
93: 141us
94: 158us
95: 879us <---------------
96: 4972us
97: 4966us
98: 4969us
99: 4969us
100: 5049us
101: 5049us
102: 5048us
103: 5048us
104: 5051us
105: 5048us
106: 5048us
107: 5048us
108: 5050us

--

Rohan Bedarkar
MS, Computer Science 

University of Chicago
rohanb at cs.uchicago.edu


 




________________________________
 From: Rohan Bedarkar <r_bedarkar at yahoo.com>
To: ZeroMQ development list <zeromq-dev at lists.zeromq.org> 
Sent: Thursday, October 25, 2012 12:32 PM
Subject: Re: [zeromq-dev] epgm performance numbers
 

Ok. Let me start with iPerf.. 

Will update this thread with results as I think it will be really useful to others as well.

 
--

Rohan Bedarkar
MS, Computer Science 

University of Chicago
rohanb at cs.uchicago.edu


 




________________________________
 From: Steven McCoy <steven.mccoy at miru.hk>
To: Rohan Bedarkar <r_bedarkar at yahoo.com>; ZeroMQ development list <zeromq-dev at lists.zeromq.org> 
Sent: Thursday, October 25, 2012 12:17 PM
Subject: Re: [zeromq-dev] epgm performance numbers
 

On 25 October 2012 13:04, Rohan Bedarkar <r_bedarkar at yahoo.com> wrote:

Unfortunately no luck. Just tried varying the multicast rates on both sides anywhere between 1000 to 10000 to even 1000000 but no change. Its a little counterintuitive that lowering rates can help especially if HWM is sufficiently high. I am using 32 bit SuSe on one side and 64 bit SuSe on the other side. 
>
>
>Are there any configs that I might have overlooked? 
>
>
>Both my repeater application and originating application are very light so I doubt these latencies are anywhere in there. If I simply switch to TCP the performance is really good.
>
>

First set a baseline with iperf.

Second try the performance tools in ZeroMQ.

Third try the performance tools in OpenPGM.

Due to the high watermark you have set in ZeroMQ it might be an issue previously raised:  application level throttling is required, although it need only be coarse grained.  The problem with high speed pushing into ZeroMQ means you saturate the memory bus queuing up everything and slow the entire system down causing performance to be significantly worse.

Reliable multicast is not easy™

-- 
Steve-o
_______________________________________________
zeromq-dev mailing list
zeromq-dev at lists.zeromq.org
http://lists.zeromq.org/mailman/listinfo/zeromq-dev



_______________________________________________
zeromq-dev mailing list
zeromq-dev at lists.zeromq.org
http://lists.zeromq.org/mailman/listinfo/zeromq-dev



_______________________________________________
zeromq-dev mailing list
zeromq-dev at lists.zeromq.org
http://lists.zeromq.org/mailman/listinfo/zeromq-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.zeromq.org/pipermail/zeromq-dev/attachments/20121029/77a44af3/attachment.htm>