[zeromq-dev] Poor performance for DEALER <-> ROUTER - ROUTER <-> DEALER in hello_world scenario .

[artemv zmq]

Ok ... Figured out!  Now answering my own very latest question -- HWM-arks
do matter.

Explaining:
If you run with defaults _everywhere_ (hwm=1K)   then  DEALER(Sender) <->
[ROUTER - ROUTER](Device) <-> DEALER(Receiver)     architecture   will give
you  ~ 600K/sec on Sender.send() and same amount on Device.send()  and same
600K  on Receiver.recv()   . But then you will have a problem in
Receiver.send()  which will be blocking almost all time and will give only
  2 ~ 5 K/sec  on sending replies back.

In case you want  not only fast  "send"  but fast  "reply"  as well (it's
my case!)   then  hwm-arks   should be set like this:

DEALER(Sender(hwm_recv=0, hwm_send=default)) <-> [ROUTER(hwm_recv=0,
hwm_send=default) - ROUTER(hwm_recv=0, hwm_send=default)](Device) <->
DEALER(Receiver(hwm_recv=0, hwm_send=default)) .   With such hwm settings I
get  almost equivalent amount of traffic in both directions:   ~ 400 K/sec!
 Sender cpu ~ 50%, Receiver  cpu ~ 30 %.  Very nice!


I much appreciate to Panu Wetterstrand! Your suggestion made everything.
Thanks again. Whenever you will be in Kyiv be sure bottle of Ballantine's
is waiting for you :)  My FB --  https://www.facebook.com/artem.vysochyn  .




2014-05-23 19:22 GMT+03:00 artemv zmq <artemv.zmq at gmail.com>:

> Hey Panu,
>
> That helped (unfortunately partially) ...  But still you made my day!
>  Many thanks.
>
> What I got now -- I can send fast all the way: from DEALER  to-device-->
> ROUTER to-> ROUTER from-device--> DEALER . This works really  fast!  Your
> suggestion given me 500K/sec.  Nice,  but that's only half of what I need,
> because   opposite flow gives ~  2 or 3 K/sec. Also I see that Sender
> process using 60% of CPU  while   Receiver using ~ 5 % of CPU, so most of
> the time Receiver  now  blocking on sending response  back.  I  debugged  a
> bit and found that bottleneck is .send() method on Receiver. So, in short:
>  the problem now -- we can send fast  but  can't   reply fast .
>
> Any ideas?
>
>
>
>
> 2014-05-23 18:06 GMT+03:00 Panu Wetterstrand <panu.wetterstrand at iki.fi>:
>
> Yes. Then there is no unnecessary poll calls.
>> 23.5.2014 17.48 kirjoitti "artemv zmq" <artemv.zmq at gmail.com>:
>>
>> Panu Wetterstrand ,
>>>
>>>
>>> Do u mean like this:
>>>
>>> ...
>>> for (; ; ) {
>>>   poller.poll(1000);
>>>   if (poller.pollin(0)) {
>>>     // block here and .recv() from socket until .recv() return null .
>>>   }
>>> }
>>> ...
>>>
>>>
>>> Is this is what  you are proposing?
>>>
>>>
>>> Thanks  for heads up /
>>>
>>>
>>>
>>>
>>> 2014-05-23 17:22 GMT+03:00 Panu Wetterstrand <panu.wetterstrand at iki.fi>:
>>>
>>>> Sorry. Event count relates to sockets so that is fine but try to read
>>>> until socket is empty and recv returns nil.
>>>> 23.5.2014 17.16 kirjoitti "Panu Wetterstrand" <panu.wetterstrand at iki.fi
>>>> >:
>>>>
>>>>  You should handle all events before calling poll again. Poll returns
>>>>> event count. Only with phone so cant write example.
>>>>> 22.5.2014 21.21 kirjoitti "artemv zmq" <artemv.zmq at gmail.com>:
>>>>>
>>>>>> I'm following advice from a guide to come with
>>>>>> as-simple-solution-as-possible before moving further, and experiencing poor
>>>>>> performance on a very simple scenario.
>>>>>>
>>>>>> My test case is to check how ROUTER/DEALER constructions are working,
>>>>>> how far they scale.  My environment:
>>>>>>  windows_7_64bit/java7/4xCPU/8g/zmq-3.2.2/jzmq-3.0.0
>>>>>>
>>>>>> Got two java processes. One process(called Sender) hosts
>>>>>> ROUTER-ROUTER device   and  forever-loop DEALER  constantly sending "hello
>>>>>> zmq world"  message.  Second process(called Receiver) is hosting  single
>>>>>> DEALER which is sort of simple worker -- receives message and sends it back.
>>>>>>
>>>>>> The whole point is measure how much messages per sec I can get at
>>>>>> this very raw scenario. Here's java code:
>>>>>>
>>>>>> Sender -- http://pastebin.com/mM7dqvSw
>>>>>> Receiver -- http://pastebin.com/dS2SLA2G
>>>>>>
>>>>>>
>>>>>> The test result ~ 40K/sec  which is much-much less than I expected.
>>>>>> Because you can't have two java processes (both taking 50% cpu)  and
>>>>>> rendering 40K  at  "hello world"  scenario..
>>>>>>
>>>>>>
>>>>>>
>>>>>> Thanks in advance for a help.
>>>>>>
>>>>>> _______________________________________________
>>>>>> 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
>>>
>>>
>> _______________________________________________
>> 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/20140523/4fa8d15c/attachment.htm>