[Swift-devel] ws-gram tests

Fri Feb 8 11:27:29 CST 2008

On Fri, 2008-02-08 at 11:19 -0600, feller at mcs.anl.gov wrote:
> Mihael,
> 
> i think i found the memory hole in GramJob.
> 100 jobs in a test of mine consumed about 23MB (constantly
> growing) before the fix and 8MB (very slowly growing) after
> the fix. The big part of that (7MB) is used right from the
> first job which may be the NotificationConsumerManager.
> Will commit that change soon to 4.0 branch and you may try
> it then.
> Are you using 4.0.x in your tests?

Yes. If there are no API changes, you can send me the jar file. I don't
have enough knowledge to selectively build WS-GRAM, nor enough disk
space to build the whole GT.

> 
> Martin
> 
> >>> >
> >>> > These are both hacks. I'm not sure I want to go there. 300K per job
> >>> is
> >>> a
> >>> > bit too much considering that swift (which has to consider many more
> >>> > things) has less than 10K overhead per job.
> >>> >
> >>>
> >>>
> >>> For my better understanding:
> >>> Do you start up your own notification consumer manager that listens for
> >>> notifications of all jobs or do you let each GramJob instance listen
> >>> for
> >>> notifications itself?
> >>> In case you listen for notifications yourself: do you store
> >>> GramJob objects or just EPR's of jobs and create GramJob objects if
> >>> needed?
> >>
> >> Excellent points. I let each GramJob instance listen for notifications
> >> itself. What I observed is that it uses only one container for that.
> >>
> >
> > Shoot! i didn't know that and thought there would be a container per
> > GramJob in that case. That's the core mysteries with notifications.
> > Anyway: I did a quick check some days ago and found that GramJob is
> > surprisingly greedy regarding memory as you said. I'll have to further
> > check what it is, but will probably not do that before 4.2 is out.
> >
> >
> >> Due to the above, a reference to the GramJob is kept anyway, regardless
> >> of whether that reference is in client code or the local container.
> >>
> >> I'll try to profile a run and see if I can spot where the problems are.
> >>
> >>>
> >>> Martin
> >>>
> >>> >>
> >>> >> The core team will be looking at improving notifications once their
> >>> >> other 4.2 deliverables are done.
> >>> >>
> >>> >> -Stu
> >>> >>
> >>> >> Begin forwarded message:
> >>> >>
> >>> >> > From: feller at mcs.anl.gov
> >>> >> > Date: February 1, 2008 9:41:05 AM CST
> >>> >> > To: "Jaime Frey" <jfrey at cs.wisc.edu>
> >>> >> > Cc: "Stuart Martin" <smartin at mcs.anl.gov>, "Terrence Martin"
> >>> >> <tmartin at physics.ucsd.edu
> >>> >> > >, "Martin Feller" <feller at mcs.anl.gov>, "charles bacon"
> >>> >> <bacon at mcs.anl.gov
> >>> >> > >, "Suchandra Thapa" <sthapa at ci.uchicago.edu>, "Rob Gardner"
> >>> >> <rwg at hep.uchicago.edu
> >>> >> > >, "Jeff Porter" <rjporter at lbl.gov>, "Alain Roy"
> >>> <roy at cs.wisc.edu>,
> >>> >> > "Todd Tannenbaum" <tannenba at cs.wisc.edu>, "Miron Livny"
> >>> >> <miron at cs.wisc.edu
> >>> >> > >
> >>> >> > Subject: Re: Condor-G WS GRAM memory usage
> >>> >> >
> >>> >> >> On Jan 31, 2008, at 6:26 PM, Jaime Frey wrote:
> >>> >> >>
> >>> >> >>> On Jan 30, 2008, at 12:25 PM, Stuart Martin wrote:
> >>> >> >>>
> >>> >> >>>> On Jan 30, 2008, at Jan 30, 11:46 AM, Jaime Frey wrote:
> >>> >> >>>>
> >>> >> >>>>> Terrence Martin's scalability testing of Condor-G with WS GRAM
> >>> >> >>>>> raised some concerns about memory usage on the client side. I
> >>> did
> >>> >> >>>>> some profiling of Condor-G's WS GRAM GAHP server, which
> >>> appeared
> >>> >> >>>>> to be the primary memory consumer. The GAHP server is a
> >>> wrapper
> >>> >> >>>>> around the java client libraries for WS GRAM.
> >>> >> >>>>>
> >>> >> >>>>> In my tests, I submitted variable numbers of jobs up to 30 at
> >>> a
> >>> >> >>>>> time. The jobs were 2-minute sleep jobs with minimal data
> >>> >> >>>>> transfer. All of the jobs overlapped in submission and
> >>> execution.
> >>> >> >>>>> Here is what I've discovered so far.
> >>> >> >>>>>
> >>> >> >>>>> Aside from the heap available to the java code, the jvm used
> >>> 117
> >>> >> >>>>> megs of non-shared memory and 74 megs of shared memory.
> >>> Condor-G
> >>> >> >>>>> creates one GAHP server for each (local uid, X509 DN) pair.
> >>> >> >>>>>
> >>> >> >>>>> The maximum jvm heap usage (as reported by the garbage
> >>> collector)
> >>> >> >>>>> was about 9 megs plus 0.9 megs per job. When the GAHP was
> >>> >> >>>>> quiescent (jobs executing, Condor-G waiting for them to
> >>> complete),
> >>> >> >>>>> heap usage was about 5 megs plus 0.6 megs per job.
> >>> >> >>>>>
> >>> >> >>>>> The only long-term memory per job that I know of in the GAHP
> >>> is
> >>> >> >>>>> for the notification sink for job status callbacks. 600kb
> >>> seems
> >>> a
> >>> >> >>>>> little high for that. Stu, could someone on Globus help us
> >>> >> >>>>> determine if we're using the notification sinks inefficiently?
> >>> >> >>>>
> >>> >> >>>> Martin just looked and for the most part, there is nothing
> >>> wrong
> >>> >> >>>> with how condor-g manages the callback sink.
> >>> >> >>>> However, one improvement that would reduce the memory used per
> >>> job
> >>> >> >>>> would be to not have a notification consumer per job.  Instead
> >>> use
> >>> >> >>>> one for all jobs.
> >>> >> >>>>
> >>> >> >>>> Also, Martin recently did some analysis on condor-g stress
> >>> tests
> >>> >> >>>> and found that notifications are building up on the in the
> >>> GRAM4
> >>> >> >>>> service container and that is causing delays which seem to be
> >>> >> >>>> causing multiple problems.  We're looking at this in a separate
> >>> >> >>>> effort with the GT Core team.  But, after this was clear,
> >>> Martin
> >>> >> >>>> re-
> >>> >> >>>> ran the condor-g test and relied on polling between condor-g
> >>> and
> >>> >> >>>> the GRAM4 service instead of notifications.  Jaime, could you
> >>> >> >>>> repeat the no-notification test and see the difference in
> >>> memory?
> >>> >> >>>> The changes would be to increase the polling frequency in
> >>> condor-g
> >>> >> >>>> and comment out the subscribe for notification.  You could also
> >>> >> >>>> comment out the notification listener call(s) too.
> >>> >> >>>
> >>> >> >>>
> >>> >> >>> I did two new sets of tests today. The first used more efficient
> >>> >> >>> callback code in the GAHP (one notification consumer rather than
> >>> one
> >>> >> >>> per job). The second disabled notifications and relied on
> >>> polling
> >>> >> >>> for job status changes.
> >>> >> >>>
> >>> >> >>> The more efficient callback code did not produce a noticeable
> >>> >> >>> reduction in memory usage.
> >>> >> >>>
> >>> >> >>> Disabling notifications did reduce memory usage. The maximum jvm
> >>> >> >>> heap usage was roughly 8 megs plus 0.5 megs per job. The minimum
> >>> >> >>> heap usage after job submission and before job completion was
> >>> about
> >>> >> >>> 4 megs + 0.1 megs per job.
> >>> >> >>
> >>> >> >>
> >>> >> >> I ran one more test with the improved callback code. This time, I
> >>> >> >> stopped storing the notification producer EPRs associated with
> >>> the
> >>> >> >> GRAM job resources. Memory usage went down markedly.
> >>> >> >>
> >>> >> >> I was told the client had to explicitly destroy these serve-side
> >>> >> >> notification producer resources when it destroys the job,
> >>> otherwise
> >>> >> >> they hang around bogging down the server. Is this still the case?
> >>> The
> >>> >> >> server can't destroy notification producers when their sources of
> >>> >> >> information are destroyed?
> >>> >> >>
> >>> >> >
> >>> >> > This reminds me of the odd fact that i had to suddenly grant much
> >>> more
> >>> >> > memory to Condor-g as soon as condor-g started storing EPRs of
> >>> >> > subscription resources to be able to destroy them eventually.
> >>> >> > Those EPR's are maybe not so tiny as they look like.
> >>> >> >
> >>> >> > For 4.0: yes, currently you'll have to store and eventually
> >>> destroy
> >>> >> > subscription resources manually to avoid heaping up persistence
> >>> data
> >>> >> > on the server-side.
> >>> >> > For 4.2: no, you won't have to store them. A job resource will
> >>> >> > destroy all subscription resources when it's destroyed.
> >>> >> >
> >>> >> > Overall i suggest to concentrate on 4.2 gram since the "container
> >>> >> > hangs in job destruction" problem won't exist anymore.
> >>> >> >
> >>> >> > Sorry, Jaime, i still can't provide you with 100% reliable 4.2
> >>> changes
> >>> >> > in Gram in 4.2. I'll do so as soon as i can. I wonder if it makes
> >>> >> > sense
> >>> >> > for us to do the 4.2-related changes in Gahp and hand it to you
> >>> for
> >>> >> > fine-tuning then?
> >>> >> >
> >>> >> > Martin
> >>> >>
> >>> >>
> >>> >>
> >>> >>
> >>> >> On Feb 8, 2008, at Feb 8, 9:19 AM, Ian Foster wrote:
> >>> >>
> >>> >> > Mihael:
> >>> >> >
> >>> >> > That's great, thanks!
> >>> >> >
> >>> >> > Ian.
> >>> >> >
> >>> >> > Mihael Hategan wrote:
> >>> >> >> I did a 1024 job run today with ws-gram.
> >>> >> >> I painted the results here:
> >>> >> >> http://www-unix.mcs.anl.gov/~hategan/s/g.html
> >>> >> >>
> >>> >> >> Seems like client memory per job is about 370k. Which is quite a
> >>> lot.
> >>> >> >> What kinda worries me is that it doesn't seem to go down after
> >>> the
> >>> >> >> jobs
> >>> >> >> are done, so maybe there's a memory leak, or maybe the garbage
> >>> >> >> collector
> >>> >> >> doesn't do any major collections. I'll need to profile this to
> >>> see
> >>> >> >> exactly what we're talking about.
> >>> >> >>
> >>> >> >> The container memory is figured by looking at the process in
> >>> /proc.
> >>> >> >> It's
> >>> >> >> total memory including shared libraries and things. But libraries
> >>> >> >> take a
> >>> >> >> fixed amount of space, so a fuzzy correlation can probably be
> >>> made.
> >>> >> >> It
> >>> >> >> looks quite similar to the amount of memory eaten on the client
> >>> side
> >>> >> >> (per job).
> >>> >> >>
> >>> >> >> CPU-load-wise, WS-GRAM behaves. There is some work during the
> >>> time
> >>> >> >> the
> >>> >> >> jobs are submitted, but the machine itself seems responsive. I
> >>> have
> >>> >> >> yet
> >>> >> >> to plot the exact submission time for each job.
> >>> >> >>
> >>> >> >> So at this point I would recommend trying ws-gram as long as
> >>> there
> >>> >> >> aren't too many jobs involved (i.e. under 4000 parallel jobs),
> >>> and
> >>> >> >> while
> >>> >> >> making sure the jvm has enough heap. More than that seems like a
> >>> >> >> gamble.
> >>> >> >>
> >>> >> >> Mihael
> >>> >> >>
> >>> >> >> _______________________________________________
> >>> >> >> Swift-devel mailing list
> >>> >> >> Swift-devel at ci.uchicago.edu
> >>> >> >> http://mail.ci.uchicago.edu/mailman/listinfo/swift-devel
> >>> >> >>
> >>> >> >>
> >>> >> >
> >>> >>
> >>> >
> >>> >
> >>>
> >>>
> >>
> >>
> >
> >
> >
> 
> 