Hi Lohit, Setting yarn.scheduler.fair.assignmultiple to false should give a more even distribution of tasks.
-Sandy On Fri, Dec 14, 2012 at 10:22 AM, lohit <[email protected]> wrote: > I should have mentioned. I was trying this on FairScheduler. > I can try CapacityScheduler and see if that works. > But this is pretty easy to reproduce as I described in the problem. > > 2012/12/14 Harsh J <[email protected]> > > > Perhaps you can try YARN-80's > > yarn.scheduler.capacity.node-locality-delay set to 1 for your > > CapacityScheduler (are you using this or the Fifo)? > > > > On Fri, Dec 14, 2012 at 1:37 AM, lohit <[email protected]> > wrote: > > > I am running hadoop-2.0.3 alpha branch. > > > Right now I am trying TestDFSIO, where I would like to have even > > > distribution of tasks. > > > As soon as I submit job RM tends to allocate as many containers as > > possible > > > from subset of nodes. > > > > > > 2012/12/13 Thomas Graves <[email protected]> > > > > > >> What kind of job are you running? What version of YARN? > > >> > > >> Tom > > >> > > >> On 12/13/12 12:24 AM, "lohit" <[email protected]> wrote: > > >> > > >> >Hi Devs, > > >> > > > >> >I am running few experiments on YARN. On a 50 node cluster I run only > > one > > >> >job. > > >> >If I kick off job with 150 tasks, I would expect on an average 3 > tasks > > in > > >> >containers per node. > > >> >But I see many nodes end up having 29-30 containers spawned and many > > >> >without any tasks at all. > > >> >Do I have to use some kind of parameter (delayed scheduling) to > achieve > > >> >even distribution? > > >> > > > >> >-- > > >> >Have a Nice Day! > > >> >Lohit > > >> > > >> > > > > > > > > > -- > > > Have a Nice Day! > > > Lohit > > > > > > > > -- > > Harsh J > > > > > > -- > Have a Nice Day! > Lohit >
