[zeromq-dev] performance & ressource tradeoff for sockets ws threads ?

Laurent Alebarde l.alebarde at free.fr
Sat Feb 15 19:41:13 CET 2014


This makes proxying CURVE quite useless, as I can instead increase the 
number of I/O threads.

Besides, in CURVE, I have introduced some delays between the different 
steps of the handcheck (part of the PARANO specifications). So here, I 
have a problem since I cannot serve hundreds of simultaneous connexions 
without introducing some huge latencies. Say I have one I/O thread, 100 
simultaneous clients, and a 100 ms handcheck, the last client sees a 10 
sec latency since the delays are queued. So here I should attribute say 
20 or 50 threads to zmq I/O and a few threads only for the workers, as I 
can use several sockets per worker.

Or I can still proxy CURVE and use different contexts in the workers, so 
that they get each their own I/O thread(s). Is it correct ? This 
solution would not reduce the number of threads: 3 threads per worker (2 
for I/O), 10 sockets per worker (max latency is 500 ms), 10 workers, 2 
for the proxy = 32 threads.

Another idea which use CURVE proxying is to manage the delays I have 
introduced, not in CURVE itself, but in the proxy. Then, the delays are 
not seen by the I/O threads, and I can get the whole application running 
with a very minimal number of threads: one for the I/O, one for the 
proxy (not blocking on the delays), and say 6 for the workers (say 100 
sockets each).

Is there some obvious strategy in terms of architecture ?




Le 14/02/2014 10:31, Pieter Hintjens a écrit :
> When you're doing encryption you will be hitting CPU limits. The work
> will happen in the background I/O thread, not your application thread.
> So start as many I/O threads as you have CPU cores. How you match
> sockets to application threads is going to be insignificant.
>
> On Thu, Feb 13, 2014 at 6:08 PM, Laurent Alebarde <l.alebarde at free.fr> wrote:
>> Hi all,
>>
>> In a server, I want to assign one socket per client (CURVE). What is the
>> best in terms of performance & ressources ? Say I want to deal with 1,000
>> simultaneous clients.
>>
>> one socket per thread with an avarage cpu load <1%, and 1,000 threads ?
>> 100 sockets per thread with an avarage cpu load near 100%, and 10 threads ?
>> anything between ?
>>
>> Some service delivery latency is acceptable.
>>
>>
>> Laurent
>>
>> _______________________________________________
>> 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/20140215/0acc1b7e/attachment.htm>


More information about the zeromq-dev mailing list