On 4/13/13 2:02 AM, Sanne Grinovero wrote:
@All, the performance problem seemed to be caused by a problem in
JGroups, which I've logged here:
https://issues.jboss.org/browse/JGRP-1617
Almost no information attached to the case :-( If it wasn't you, Sanne,
I'd outright reject the case ...
The MessageDispatcher will *not* wait until the timeout kicks in, it'll
return as soon as it has acks from all members of the target set. This
works and is covered with a bunch of unit tests, so a regression would
have been caught immediately.
I attached a test program to JGRP-1617 which shows that this feature
works correctly.
Of course, if you lose an ack (e.g. due to a maxed out incoming / OOB
thread pool), the unicast protocol will have to retransmit the ack until
it has been received. Depending on the unicast protocol you use, this
will be immediate (UNICAST, UNICAST3), or based on a stability interval
(UNICAST2).
For the record, the first operation was indeed triggering some lazy
initialization of indexes, which in turn would trigger a Lucene
Directory being started, triggering 3 Cache starts which in turn would
trigger 6 state transfer processes: so indeed the first operation
would not be exactly "cheap" performance wise, still this would
complete in about 120 milliseconds.
This sounds very low for the work you describe above. I don't think 6
state transfers can be completed in 120ms, unless they're async (but
then that means they're not done when you return). Also, cache starts
(wrt JGroups) will definitely take more than a few seconds if you're the
first cluster node...
Not being sure about the options of depending to a newer JGroups
release or the complexity of a fix, I'll implement a workaround in
HSearch in the scope of HSEARCH-1296.
If you add more information to JGRP-1617, I'll take a look. This would
be a critical bug in JGroups *if* you can prove that the
MessageDispatcher always runs into the timeout (I don't think you can
though !).
As a lesson learned, I think we need to polish some of our TRACE
level
messaged to include the cache name: to resolve this we had not just
many threads and components but also 4 of them where using JGroups
(interleaving messages of all sorts) and 9 different caches where
involved for each simple write operation in CD: made it interesting to
figure what was going on!
Yes, that would help. In JGroups, I usually log the cluster address of
the thread that's writing to the log, so I can differentiate between
different clusters on the same host.
Also I'm wondering how hard it would be to
have a log parser which converts my 10GB of text log from today in a
graphical sequence diagram.
Yes, something like wireshark "follow TCP" feature would be very helpful !
--
Bela Ban, JGroups lead (
http://www.jgroups.org)