Although some control sockets may not require HWM, I follow
the rule "HWM for any socket". Besides this I have some TTL-alike
marks to drop old messages out.
It is a design decision: sometimes the right thing to do is to drop a message out,
sometimes to use blocking approach and wait.
Coming back to the issue, it must be clear if a process' memory footprint goes
high the process is too slow and perhaps a sort of balancing
should be used. Situation is different is case of blocking sockets and/or
inproc communications.
16.05.2011, 17:54, "Andrew Hume" <[email protected]>:
how can i tell if it was hit?generally, the process links don't have hwm set, but some do.is your advice to set them on all links?On May 16, 2011, at 6:51 AM, Ian Barber wrote:On Mon, May 16, 2011 at 2:40 PM, Andrew Hume <[email protected]> wrote:_______________________________________________i can increase the number of processes in various parts of the processing graph,but how can i effectively figure out what to increase? ordinarily, i would look atwhich processes have increasing input queue lengths. but 0mq doesn't do that.all i can measure is the memory footprint which starts increasing, sometimesalarmingly quickly, but mostly steadily. i can't tell if the memory usage is fromfragmentation, or an input queue, or output queue.
There was quite a discussion about this sort of monitoring at the unconf, so there may be some more library level functionality coming that helps. That said, have you tried lowering the HWM to a minimum value, and collecting statistics on how often that is hit? If you're experiencing congestion at a certain point, the queue size should grow beyond the minimums fairly quickly I would guess.
Ian
zeromq-dev mailing list
[email protected]
http://lists.zeromq.org/mailman/listinfo/zeromq-dev
------------------Andrew Hume (best -> Telework) +1 623-551-2845AT&T Labs - Research; member of USENIX and LOPSA_______________________________________________
zeromq-dev mailing list
[email protected]
http://lists.zeromq.org/mailman/listinfo/zeromq-dev
--
Best regards,
Ilja Golshtein.
Best regards,
Ilja Golshtein.
_______________________________________________ zeromq-dev mailing list [email protected] http://lists.zeromq.org/mailman/listinfo/zeromq-dev
