[zeromq-dev] publish/subscribe and the "fast producer" problem

[Marcin Romaszewicz]

Yeah, you need a way to throttle back the publisher. Have the many nodes
tell the publisher to slow down if they detect too high of a drop rate.
Your publisher can then decide whether to throttle for the slowest client,
or for the fastest client. Also, pub/sub may not be the right way to do
this for 1:many if the value of 'many' is very high, because each message
you send will be sent 1:1 to each subscriber, slowing them all down. Some
kind of tree based fanout may work better.

-- Marcin

On Fri, Mar 3, 2017 at 12:49 PM, Justin Karneges <justin at karneges.com>
wrote:

> And what if there are no subscribers? Publishing should block?
>
> It sounds like your goal is some kind of sensible flow control (I'm not
> sure what else an "at least 1 subscriber can read" use-case really is). My
> advice is to publish at a fixed rate, for example by sleeping between
> publishes.
>
> I do wish ZMQ_RATE worked for all PUB sockets and not just PGM-based.
>
> On Fri, Mar 3, 2017, at 12:31 PM, Francesco wrote:
>
> Hi all,
>
> I'm new to ZeroMQ but so far all my experiments with it have been quite
> positive, so thanks for your great work!
>
> However, one thing that I cannot really sort out is the following: I have
> written a small sample program where I create a ZMQ_PUB socket and I
> continuously zmq_msg_send() messages in it, never sleeping. This is
> simulating what I will do in a larger program where I plan to use ZeroMQ.
> Such program is a massively-parallel utility that needs to send in a fanout
> fashion (1 to many) several short messages per second (up to say 1million
> messages / second).
>
> In my scenario I don't care about subscribers joining late or eventually
> loosing some messages (if the network is slow or the subscriber itself is
> too slow). I do care however about detecting such conditions where messages
> are dropped.
>
> My problem is the "fast producer" one: this sample program shows that the
> zmq_msg_send() never returns an error; doing some math such my little
> utility says that is publishing data at rates up to 400Gbps...  the only
> problem is that the NIC on the computer is a 1Gbps NIC. Of course all
> subscribers report missing 99% of messages (I put a sequence number in the
> messages I send).
>
> This problem does not appear to be new, some interesting references I
> found are:
>
>    http://grokbase.com/t/zeromq/zeromq-dev/11ca11s9b9/pub-sub-
> pattern-rate-control-and-backpressure            (Dec 2011)
>    http://grokbase.com/t/zeromq/zeromq-dev/12797gy703/notify-
> send-er-that-theyve-hit-the-high-water-mark      (Jul 2012)
>    http://grokbase.com/t/zeromq/zeromq-dev/129n2e2sx5/high-
> water-mark-notification-for-publisher                   (Sep 2012)
>
> I experimented a little bit with ZeroMQ and found that:
>   - ZMQ_RATE is not doing anything
>   - ZMQ_XPUB_NODROP set to 1 works BUT basically ties the publisher to the
> SLOWEST subscriber: from my experiments it looks like (despite the
> documentation) the publisher now blocks every time there is just 1
> subscriber queue that has hit the HWM
>
> as written in some of these posts what I would like to have is a NODROP
> socket option that allows my publisher to send data as fast as the FASTEST
> subscriber can handle.
>
> So here's my question: is there any way to achieve the above, i.e. block
> the publisher if ALL subscriber queues hit their HWM ?
>
>
> Thanks a lot for any hint!
>
> Francesco
>
>
>
>
>
> *_______________________________________________*
> zeromq-dev mailing list
> zeromq-dev at lists.zeromq.org
> https://lists.zeromq.org/mailman/listinfo/zeromq-dev
>
>
>
> _______________________________________________
> zeromq-dev mailing list
> zeromq-dev at lists.zeromq.org
> https://lists.zeromq.org/mailman/listinfo/zeromq-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.zeromq.org/pipermail/zeromq-dev/attachments/20170303/71b70fbe/attachment.htm>