Alright, thanks, we're already busy rolling out the config for the capacity scheduler. Still, interesting behavior. The fifo scheduler looks at the load on the cores? Seems unnecessary, the kernel is quite good at context switching.
Evert ________________________________________ From: Harsh J [ha...@cloudera.com] Sent: Tuesday, August 14, 2012 3:12 PM To: user@hadoop.apache.org Subject: Re: Pending reducers I guess this is the regular behavior of the default FIFO task scheduler. It takes into account the reducer load and that may be why it refused to schedule the rest up immediately. You may have better luck using either Fair or Capacity schedulers. On Tue, Aug 14, 2012 at 5:56 PM, Evert Lammerts <evert.lamme...@sara.nl> wrote: >> whats the memory/cpu stats on the machines ? are they exhausted > > No, they're not. The nodes themselves have more than enough memory available, > and the load on the cores sits between 0.8 and 0.9. > > Is current load in terms other than available slots even taken into account > in the default scheduler? That would surprise me, actually... But it might > explain this behavior. > > Evert > >> >> On Tue, Aug 14, 2012 at 5:20 PM, Evert Lammerts >> <evert.lamme...@sara.nl> wrote: >> >> reducers of multiple jobs do run con-currently as long as they have >> >> the resources available. >> > >> > Yep, and that's what's not happening in my situation. 528 reduce >> slots, 400 taken by one job, 26 of another job remain in pending state. >> What could explain this behavior? >> > >> > Evert >> > >> >> >> >> If you want to limit someone overtaking the cluster, then you can >> >> create different job queues and assign quota to each queue. You also >> >> have the flexibility of allocating max quota per user in a queue as >> >> well. >> >> >> >> >> >> >> >> On Tue, Aug 14, 2012 at 4:09 PM, Evert Lammerts >> >> <evert.lamme...@sara.nl> wrote: >> >> > Hi list, >> >> > >> >> > I have a cluster running Hadoop 0.20.205 with Kerberos enabled, >> >> exposing 528 map slots and 528 reduce slots. Currently somebody is >> >> running a NORMAL priority job with 7 mappers and 400 reducers. The >> >> mappers have finished and the system is processing the reducers. >> >> Another user is running a NORMAL priority job with 1 mapper and 26 >> >> reducers. The mapper has finished, but the reducers won't come out >> of >> >> "pending" state. There are no other jobs running right now. We've >> not >> >> yet installed a different scheduler, so right now the system is >> using >> >> the default scheduler. How can this behavior be explained? I see >> >> mappers of multiple jobs run concurrently, and I *thought* I've seen >> >> reducers of multiple jobs run concurrently, but I'm not completely >> >> sure. Any idea? >> >> > >> >> > Evert >> >> >> >> >> >> >> >> -- >> >> Nitin Pawar >> >> >> >> -- >> Nitin Pawar -- Harsh J