[zeromq-dev] Windows throughput perfomance. Once again.

Геннадий Казачёк gena.kazachek at gmail.com
Wed Jun 1 08:02:01 CEST 2011


---------- Пересланное сообщение ----------
От кого: Геннадий Казачёк <gena.kazachek at gmail.com>
Дата: 1 июня 2011 г. 10:01
Тема: Re: [zeromq-dev] Windows throughput perfomance. Once again.
Кому: Martin Sustrik <sustrik at 250bpm.com>




31 мая 2011 г. 16:50 пользователь Martin Sustrik <sustrik at 250bpm.com>написал:

On 05/31/2011 01:15 PM, Геннадий Казачёк wrote:
>
>  10000 packets
>> zmq :: mailbox_t :: recv     Hit count - 126.
>>
>> 50000 packets
>> zmq :: mailbox_t :: recv     Hit count - 611.
>>
>> Linear dependency. I guess It's a bunch of small equal waits.
>>
>
> Ok. The next thing would be to why each mailbox_t::recv takes ~1 second.
>
> The function doesn't do much, there's no way it could work for 1 second, so
> it's basically being asleep there. It can, for example, be waiting for a
> message because sender blocks for 1 sec at times. Etc
>

> Martin
>

It's a little tricky. If I make two consecutive tests I, can get very
different results. For example 3Mbit/s and 10 Mbit/s. They are both very
low, but one is 3 times greater than other, so the value of 1 second in some
function is not mandatory. In other way, when I make such a test on Linux
--> Linux it takes less than a second to finish. So, even if sleep time in
zmq :: mailbox_t :: recv will be nullified somehow, remaining more then a
hundred seconds in zmq :: select_t :: loop are a way too much. So, I think
this times are result of some other cause. For example (but absolutely not
for sure) it may be some effects with buffer overflowing on the receiver
side, when TCP denies any sending until receiver sort out.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.zeromq.org/pipermail/zeromq-dev/attachments/20110601/d760ba55/attachment.htm>


More information about the zeromq-dev mailing list