Garry, Thanks for the reply. My test wasn't very scientific. I scheduled 16 tasks at once on a test cluster of two nodes and when I looked at the ones running, one machine was maxed out (memory-wise) with 10 containers while the other had only 3 and was adding the remaining 3 tasks. It seemed the scheduling algorithm was to fill one first then the other.
Thanks, Roger On Fri, May 29, 2015 at 4:42 PM, Garry Turkington < g.turking...@improvedigital.com> wrote: > Hi Roger, > > I've seen unbalanced container assignment across hosts but never one being > maxed out before any others get any containers. So I'd look to the YARN > config to start with. > > I believe though there will be a risk of this type of thing until YARN > implements anti-affinity: > > https://issues.apache.org/jira/browse/YARN-1042 > > But as said above, maxing out a host at a time does sound odd to me. > > Garry > > -----Original Message----- > From: Roger Hoover [mailto:roger.hoo...@gmail.com] > Sent: 29 May 2015 23:09 > To: dev@samza.apache.org > Subject: Yarn scheduling > > Hi, > > I notice that when YARN schedules my jobs, it loads up one machine > completely before scheduling on the next. I'm using Capacity Scheduler > with a default config. > > Is there a way to make it "round-robin" among the available machines? > > Thanks, > > Roger > > ----- > No virus found in this message. > Checked by AVG - www.avg.com > Version: 2014.0.4800 / Virus Database: 4311/9897 - Release Date: 05/29/15 >
