Selina, Thanks for the log files. The agent log shows "Node Memory Total:251392 MB" which looks like a healthy size. Is this value correct? Just in case, what OS are you running on?
Unfortunately the RSS size for the JD and JP processes is not being shown in the agent log in that version of DUCC. They should be shown on the job details page of ducc-mon. The agent log does confirm that cgroups are not enabled, which should eliminate the possibility that the JD was swapping. That leaves me puzzled about JD behavior. The need for a JD sending references to input data rather than the data itself is to try to avoid making the JD a bottleneck when processing is scaled out. Not yet clear that the current CR is the cause of erratic behavior. I attempted to replicate your job here but got stuck on UMLS authentication and am now waiting for the approval to use UMLS. DUCC's rogue detection is intended for machines that are fully managed by DUCC. The default properties includes a small number of processes, like ssh and bash, which are useful to ignore. All UIDs below a specified threshold are also ignored. Certainly OK to customize to ignore specific users or process names, but remember that DUCC will attempt to utilize all memory that was not taken by system users when the agent started. Unexpected memory use by other processes can lead to over-committing system memory. Below are the lines to modify as desired, add to the file site.ducc.properties, and then restart DUCC. The site file facilitates migration to new DUCC versions. # max UID reserved by OS. This is used to detect rogue processes and to report # available memory on a node. ducc.agent.node.metrics.sys.gid.max=500 # exclude the following user ids while detecting rogue processes ducc.agent.rogue.process.user.exclusion.filter= #exclude the following processes while detecting rogue processes ducc.agent.rogue.process.exclusion.filter=sshd:,-bash,-sh,/bin/sh,/bin/bash,grep,ps Regards, Eddie On Fri, Mar 11, 2016 at 1:38 PM, Selina Chu <selina....@gmail.com> wrote: > Hi Eddie, > > Thanks for the pointer about not putting the analytic pipeline in the JD > driver. It seems like we’ve misunderstood the use of it. We’ll look into > modifying it so that the JD driver contains only the collection reader > component. Hopefully cTAKES will let us do so. > > As suggested, I restarted DUCC and ran the same job once. The agent.log > file is quite big. So I’ve placed it in a repo, along with others related > logs in here: > https://github.com/selinachu/Templogs/tree/master/NewLogs_Mar11 > > I noticed that the agent log indicated many rogue processes. Would it be > helpful to modify the settings in ducc.properties to clean up these > processes? > > Thanks again for your help. > > Cheers, > Selina > > > On Thu, Mar 10, 2016 at 10:30 AM, Eddie Epstein <eaepst...@gmail.com> > wrote: > > > Hi, > > > > DUCC has some logfiles that show more details of the machine and the job > > which would allow us to answer your questions about machine physical > > resources. These are located in $DUCC_HOME/logs, and in particular the > > agent log would be very helpful. The logfile name is {machine > > name}.{domain}.agent.log > > Please restart ducc so we can see the log from agent startup thru running > > the job one time. > > > > As for the JD memory requirement, the JD driver should not contain any of > > the analytic pipeline. Its purpose is normally to send a reference to the > > input data to the Job Processes which will read the input data, process > it > > and write results. (This is described at > > http://uima.apache.org/d/uima-ducc-2.0.0/duccbook.html#x1-1600008.1 ) > > > > It should be possible for you to take *just* the collection reader > > component from the cTAKES pipeline and use that for the JobDriver. > > Hopefully this would need much less than Xmx400, > > > > Regards, > > Eddie > > > > > > On Thu, Mar 10, 2016 at 12:07 PM, Selina Chu <selina....@gmail.com> > wrote: > > > > > Hi Eddie, > > > > > > Thanks so much for taking the time to look at my issue and for your > > reply. > > > > > > The reason I had to increase the heap size for the JD is because I'm > > > running cTAKES (http://ctakes.apache.org/) with DUCC. The increased > > heap > > > size is to accommodate loading all the models from cTAKES into memory. > > > Before, when I didn't increase the memory size, DUCC would cancel the > > > driver and ends. cTAKES would return back the error of > > > "java.lang.OutOfMemoryError: Java heap space”. > > > > > > Would you say that this problem is mainly a limitation of my physical > > > memory and processes that are running on my computer or can it be > > adjusted > > > in DUCC, like making parameter adjustments so I can use an increased > heap > > > size or maybe a way to pre-allocate enough memory to be used by DUCC? > > > > > > Thanks again, > > > Selina > > > > > > > > > On Wed, Mar 9, 2016 at 7:35 PM, Eddie Epstein <eaepst...@gmail.com> > > wrote: > > > > > > > Hi Selina, > > > > > > > > I suspect that the problem is due to the following job parameter: > > > > driver_jvm_args -Xmx4g > > > > > > > > This would certainly be true if cgroups have been configured on for > > DUCC. > > > > The default cgroup size for a JD is 450MB, so specifying an Xmx of > 4GB > > > can > > > > cause the JVM to spill into swap space and cause erratic behavior. > > > > > > > > Comparing a "fast" job (96) vs "slow" job (97), the time to process > the > > > > single work item was 8 sec vs 9 sec: > > > > 09 Mar 2016 08:46:08,556 INFO JobDriverHelper - T[20] summarize > > > > workitem statistics [sec] avg=8.14 min=8.14 max=8.14 stddev=.00 > > > > vs > > > > 09 Mar 2016 08:56:46,583 INFO JobDriverHelper - T[19] summarize > > > > workitem statistics [sec] avg=9.41 min=9.41 max=9.41 stddev=.00 > > > > > > > > The extra delays between the two jobs appear associated with the Job > > > > Driver. > > > > > > > > Was there some reason you specified heap size for the JD? The default > > JD > > > > heap size is Xmx400m. > > > > > > > > Regards, > > > > Eddie > > > > > > > > > > > > > > > > On Wed, Mar 9, 2016 at 2:41 PM, Selina Chu <selina....@gmail.com> > > wrote: > > > > > > > > > Hi > > > > > > > > > > I’m kind of new to DUCC and this forum. I was hoping to see if > > someone > > > > > could give me some insights as to why DUCC is behaving strangely > and > > a > > > > bit > > > > > unstable. > > > > > > > > > > So what I'm trying to do is: I’m using DUCC to process a cTAKES > job. > > > > > Currently DUCC is just using a single node. DUCC seems to act > > randomly > > > > in > > > > > processing the jobs, varying between 4.5 minutes to 23 minutes, > and I > > > > > wasn’t running anything else that is CPU intensive. When I don’t > use > > > DUCC > > > > > and use cTAKES alone, the times for processing are pretty > consistent. > > > > > > > > > > To demonstrate this strange behavior in DUCC, I submitted the exact > > > same > > > > > job 10 times in a row (job ID 95-104), without modification to the > > > > > settings. > > > > > The duration for finishing each of the jobs are: 4:41, 4:43, 12:48, > > > 8:41, > > > > > 5:24, 4:38, 7:07, 23:08, 8:08, 20:37 (canceled by system). The > first > > 9 > > > > jobs > > > > > were completed and the last one got canceled. Even before the last > > > job, > > > > > the first 9 jobs were varying in duration times. > > > > > After restarting DUCC a couple of times and resetting it, I > submitted > > > the > > > > > same job (job ID 110), that job was completed without a problem > (long > > > > > processing time) > > > > > > > > > > I noticed that when a job takes a long time to finish, past 5 > > minutes, > > > it > > > > > seemed to be stuck at the “initializing” and “completing” states > for > > > the > > > > > longest. > > > > > > > > > > It seems like DUCC is doing something randomly. I tried examining > > the > > > > log > > > > > files, but they are all similar, except for the time between each > > > state. > > > > > (I’ve also placed the related logs and job file in a repo > > > > > https://github.com/selinachu/Templogs, in case anyone is > interested > > in > > > > > examining them.) > > > > > > > > > > I’m baffled with the random behaviors from DUCC. I was hoping maybe > > > > someone > > > > > could clarify this more for me. > > > > > > > > > > After completing a job, what does DUCC do? Does it save something > in > > > > > memory, which carries over to the next job, which probably relates > to > > > the > > > > > initialization process? Are there some parameter settings that > might > > > > > alleviate this type of behavior? > > > > > > > > > > I would appreciate any insight. Thanks in advance for your help. > > > > > > > > > > > > > > > Cheers, > > > > > Selina Chu > > > > > > > > > > > > > > >
