Re: DUCC doesn't use all available machines
On Sat, Nov 29, 2014 at 4:46 PM, Simon Hafner reactorm...@gmail.com wrote: I've thrown some numbers at it (doubling each) and it's running at comfortable 125 procs. However, at about 6.1k of 6.5k items, the procs drop down to 30. 125 processes at 8 threads each = 1000 active pipelines. How CPU cores are these 1000 pipelines running on?
Re: DUCC doesn't use all available machines
2014-11-30 7:25 GMT-06:00 Eddie Epstein eaepst...@gmail.com: On Sat, Nov 29, 2014 at 4:46 PM, Simon Hafner reactorm...@gmail.com wrote: I've thrown some numbers at it (doubling each) and it's running at comfortable 125 procs. However, at about 6.1k of 6.5k items, the procs drop down to 30. 125 processes at 8 threads each = 1000 active pipelines. How CPU cores are these 1000 pipelines running on? Only 60.
Re: DUCC doesn't use all available machines
Now you are hitting a limit configured in ducc.properties: # Max number of work-item CASes for each job ducc.threads.limit = 500 62 job process * 8 threads per process = 496 max concurrent work items. This was put in to limit the memory required by the job driver. This value can probably be pushed up in the range of 700-800 before the job driver will go OOM. There are configuration parameters to increase JD memory: # Memory size in MB allocated for each JD ducc.jd.share.quantum = 450 # JD max heap size. Should be smaller than the JD share quantum ducc.driver.jvm.args = -Xmx400M -DUimaAsCasTracking DUCC would have to be restarted for the JD size parameters to take effect. One of the current DUCC development items is to significantly reduce the memory needed per work item, and raise the default limit for concurrent work items by two or three orders of magnitude. On Fri, Nov 28, 2014 at 6:40 PM, Simon Hafner reactorm...@gmail.com wrote: I've put the fudge to 12000, and it jumped immediately to 62 procs. However, it doesn't spawn new ones even though it has about 6k items left and it doesn't spawn more procs. 2014-11-17 15:30 GMT-06:00 Jim Challenger chall...@gmail.com: It is also possible that RM prediction has decided that additional processes are not needed. It appears that there were likely 64 work items dispatched, plus the 6 completed, leaving only 30 that were idle. If these work items appeared to be completing quickly, the RM would decide that scale-up would be wasteful and not do it. Very gory details if you're interested: The time to start a new processes is measured by the RM based on the observed initialization time of the processes plus an estimate of how long it would take to get a new process actually running. A fudge-factor is added on top of this because in a large operation it is wasteful to start processes (with associated preemptions) that only end up doing a few work tems. All is subjective and configurable. The average time-per-work item is also reported to the RM. The RM then looks at the number of work items remaining, and the estimated time needed to processes this work based on the above, and if it determines that the job will be completed before new processes can be scaled up and initialized, it does not scale up. For short jobs, this can be a bit inaccurate, but those jobs are short :) For longer jobs, the time-per-work-item becomes increasingly accurate so the RM prediction tends to improve and ramp-up WILL occur if the work-item time turns out to be larger than originally thought. (Our experience is that work-item times are mostly uniform with occasional outliers, but the prediction seems to work well). Relevant configuration parameters in ducc.properties: # Predict when a job will end and avoid expanding if not needed. Set to false to disable prediction. ducc.rm.prediction = true # Add this fudge factor (milliseconds) to the expansion target when using prediction ducc.rm.prediction.fudge = 12 You can observe this in the rm log, see the example below. I'm preparing a guide to this log; for now, the net of these two log lines is: the projection for the job in question (job 208927) is that 16 processes are needed to complete this job, even though the job could use 20 processes at full expanseion - the BaseCap - so a max of 16 will be scheduled for it, subject to fair-share constraint. 17 Nov 2014 15:07:38,880 INFO RM.RmJob - */getPrjCap/* 208927 bobuser O 2 T 343171 NTh 128 TI 143171 TR 6748.601431980907 R 1.8967e-02 QR 5043 P 6509 F 0 ST 1416254363603*/return 16/* 17 Nov 2014 15:07:38,880 INFO RM.RmJob - */initJobCap/* 208927 bobuser O 2 */Base cap:/* 20 Expected future cap: 16 potential cap 16 actual cap 16 Jim On 11/17/14, 3:44 PM, Eddie Epstein wrote: DuccRawTextSpec.job specifies that each job process (JP) run 8 analytic pipeline threads. So for this job with 100 work items, no more than 13 JPs would ever be started. After successful initialization of the first JP, DUCC begins scaling up the number of JPs using doubling. During JP scale up the scheduler monitors the work item completion rate, compares that with the JP initialization time, and stops scaling up JPs when starting more JPs will not make the job run any faster. Of course JP scale up is also limited by the job's fair share of resources relative to total resources available for all preemptable jobs. To see more JPs, increase the number and/or size of the input text files, or decrease the number of pipeline threads per JP. Note that it can be counter productive to run too many pipeline threads per machine. Assuming analytic threads are 100% CPU bound, running more threads than real cores will often slow down the overall document processing rate. On Mon, Nov 17, 2014 at 6:48 AM, Simon Hafner
Re: DUCC doesn't use all available machines
It is also possible that RM prediction has decided that additional processes are not needed. It appears that there were likely 64 work items dispatched, plus the 6 completed, leaving only 30 that were idle. If these work items appeared to be completing quickly, the RM would decide that scale-up would be wasteful and not do it. Very gory details if you're interested: The time to start a new processes is measured by the RM based on the observed initialization time of the processes plus an estimate of how long it would take to get a new process actually running. A fudge-factor is added on top of this because in a large operation it is wasteful to start processes (with associated preemptions) that only end up doing a few work tems. All is subjective and configurable. The average time-per-work item is also reported to the RM. The RM then looks at the number of work items remaining, and the estimated time needed to processes this work based on the above, and if it determines that the job will be completed before new processes can be scaled up and initialized, it does not scale up. For short jobs, this can be a bit inaccurate, but those jobs are short :) For longer jobs, the time-per-work-item becomes increasingly accurate so the RM prediction tends to improve and ramp-up WILL occur if the work-item time turns out to be larger than originally thought. (Our experience is that work-item times are mostly uniform with occasional outliers, but the prediction seems to work well). Relevant configuration parameters in ducc.properties: # Predict when a job will end and avoid expanding if not needed. Set to false to disable prediction. ducc.rm.prediction = true # Add this fudge factor (milliseconds) to the expansion target when using prediction ducc.rm.prediction.fudge = 12 You can observe this in the rm log, see the example below. I'm preparing a guide to this log; for now, the net of these two log lines is: the projection for the job in question (job 208927) is that 16 processes are needed to complete this job, even though the job could use 20 processes at full expanseion - the BaseCap - so a max of 16 will be scheduled for it, subject to fair-share constraint. 17 Nov 2014 15:07:38,880 INFO RM.RmJob - */getPrjCap/* 208927 bobuser O 2 T 343171 NTh 128 TI 143171 TR 6748.601431980907 R 1.8967e-02 QR 5043 P 6509 F 0 ST 1416254363603*/return 16/* 17 Nov 2014 15:07:38,880 INFO RM.RmJob - */initJobCap/* 208927 bobuser O 2 */Base cap:/* 20 Expected future cap: 16 potential cap 16 actual cap 16 Jim On 11/17/14, 3:44 PM, Eddie Epstein wrote: DuccRawTextSpec.job specifies that each job process (JP) run 8 analytic pipeline threads. So for this job with 100 work items, no more than 13 JPs would ever be started. After successful initialization of the first JP, DUCC begins scaling up the number of JPs using doubling. During JP scale up the scheduler monitors the work item completion rate, compares that with the JP initialization time, and stops scaling up JPs when starting more JPs will not make the job run any faster. Of course JP scale up is also limited by the job's fair share of resources relative to total resources available for all preemptable jobs. To see more JPs, increase the number and/or size of the input text files, or decrease the number of pipeline threads per JP. Note that it can be counter productive to run too many pipeline threads per machine. Assuming analytic threads are 100% CPU bound, running more threads than real cores will often slow down the overall document processing rate. On Mon, Nov 17, 2014 at 6:48 AM, Simon Hafner reactorm...@gmail.com wrote: I fired the DuccRawTextSpec.job on a cluster consisting of three machines, with 100 documents. The scheduler only runs the processes on two machines instead of all three. Can I mess with a few config variables to make it use all three? id:22 state:Running total:100 done:0 error:0 retry:0 procs:1 id:22 state:Running total:100 done:0 error:0 retry:0 procs:2 id:22 state:Running total:100 done:0 error:0 retry:0 procs:4 id:22 state:Running total:100 done:1 error:0 retry:0 procs:8 id:22 state:Running total:100 done:6 error:0 retry:0 procs:8
