On 25/07/18 04:52, David Mathog wrote:
One possibility is that at the "leading" edge the first job that reads a section of data will do so slowly, while later jobs will take the same data out of cache. That will lead to a "peloton" sort of effect, where the leader is slowed and the followers accelerated. iostat didn't show very much disk IO though.
I have to admit that was my first thought too. I also started to speculate about power saving but I couldn't see a way there for later jobs to catch up enough. One fun thing would be to turn HT off and set the scheduler to run 20 jobs at a time and see if it still happens then. Perhaps running this step with "perf record" to try and capture profile data and then look to see if you can spot differences across all the runs? Not sure if there are scripts to do that, or how easy it would be to rig up (plus of course the extra I/O of recording the traces will perturb the system). A very interesting problem! All the best, Chris -- Chris Samuel : http://www.csamuel.org/ : Melbourne, VIC _______________________________________________ Beowulf mailing list, Beowulf@beowulf.org sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf