[zeromq-dev] zmq performance over epgm

Steven McCoy steven.mccoy at miru.hk
Wed May 14 18:37:54 CEST 2014


On 9 May 2014 07:37, <space3001 at mail.bg> wrote:

> I am trying to send 500000 bytes of data via multicast using epgm protocol
> and the zmq framework. It takes around 45 seconds for the whole data to
> arrive at the receiver. HWM is set to 65000000, ZMQ_RATE is 100 - for
> some reason with higher rate more time is needed. Send and receive BUF are
> 20000000. When the data is being sent 100% of cpu is used by the
> publisher. If I am interpreting the data given by Wireshark right, there is
> no dropping and not too many retransmits.
>
Surprising but here are the general points:

   - TCP is accelerated and UDP is not thus there are significant CPU
   overheads in transmission from segmentation to checksumming.
   - Higher bandwidth rates can be counterproductive when the underlying
   system cannot achieve a stable throughput at that rate, this occurs
   predominantly with Windows.
   - The CPU may end up spinning with insufficiently stable timers but high
   resolution PGM parameters.

Generally multicast is preferred for larger scale out whilst TCP is pretty
good these days for moderate usage.

-- 
Steve-o
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.zeromq.org/pipermail/zeromq-dev/attachments/20140514/142c6eec/attachment.htm>


More information about the zeromq-dev mailing list