Hi Greg, It would be very interesting to do a profiling of the job master to see what it mostly spends time on. Did you run your experiments with 0.9.X or the 0.10-SNAPSHOT? Would be interesting to know if there is a regression.
Best, Max On Wed, Oct 21, 2015 at 10:08 AM, Till Rohrmann <trohrm...@apache.org> wrote: > Hi Greg, > > there is no official guide for running Flink on large clusters. As far as I > know, the cluster we used for the matrix factorization was the largest > cluster we've run a serious job on. Thus, it would be highly interesting to > understand what made the JobManager to slow down. At some point, though, > this should happen since the JobManager always stays a single instance. Do > you have by chance access to the JobManager log file? This might be helpful. > > Thanks for your help, > Till > > On Tue, Oct 20, 2015 at 11:06 PM, Greg Hogan <c...@greghogan.com> wrote: > >> Is there guidance for configuring Flink on large clusters? I have recently >> been working to benchmark some algorithms on and test AWS. I had no issues >> running on a 16 node cluster but when moving to 64 nodes the JobManager >> struggled mightily. It did not look to be parallelizing its workload. I was >> in the process of modifying my code to reduce the parallelism of earlier, >> smaller operations when I lost the cluster due to a spot price increase. >> >> The instances were c3.8xlarge and in the larger cluster one instance hosted >> the JobManager so the parallelism was 63 * 32 = 2016. The small cluster had >> parallelism of 512. >> >> I have seen the blog posts describing the performance of 640 core clusters >> on GCE. Is this a known limitation or can Flink scale much further? >> >> >> http://data-artisans.com/computing-recommendations-at-extreme-scale-with-apache-flink/ >> >> >> http://data-artisans.com/how-to-factorize-a-700-gb-matrix-with-apache-flink/ >> >> Thanks, >> Greg >>