Gregory, I've been given a tip by one of the ceph user list members on tuning values and data migration and cluster IO. I had an issues twice already where my vms would simply loose IO and crash while the cluster is being optimised for the new tunables. The recommendations were to upgrade the cluster and migrate all of your vms at least once so that the migrated vms are launched using the new version of ceph libraries. Once this is done you should be okay with data movement and your vms should have IO. Andrei ----- Original Message ----- > From: "Gregory Farnum" <gfarnum@xxxxxxxxxx> > To: "Goncalo Borges" <goncalo.borges@xxxxxxxxxxxxx> > Cc: "ceph-users" <ceph-users@xxxxxxxx> > Sent: Tuesday, 9 August, 2016 02:34:11 > Subject: Re: Advice on migrating from legacy tunables to Jewel tunables. > On Mon, Aug 8, 2016 at 5:14 PM, Goncalo Borges > <goncalo.borges@xxxxxxxxxxxxx> wrote: >> Thanks for replying Greg. >> >> I am trying to figure oout what parameters should I tune to mitigate the >> impact of the data movement. For now, I've set >> >> osd max backfills = 1 >> >> Are there others you think we should set? >> >> What do you reckon? > > That is generally the big one, but I think you'll need advice from > people who actually run clusters to see if there's anything more > that's useful. :) > -Greg > >> >> Cheers >> >> Goncalo >> >> >> On 08/09/2016 09:26 AM, Gregory Farnum wrote: >>> >>> On Thu, Aug 4, 2016 at 8:57 PM, Goncalo Borges >>> <goncalo.borges@xxxxxxxxxxxxx> wrote: >>>> >>>> Dear cephers... >>>> >>>> I am looking for some advice on migrating from legacy tunables to Jewel >>>> tunables. >>>> >>>> What would be the best strategy? >>>> >>>> 1) A step by step approach? >>>> - starting with the transition from bobtail to firefly (and, in this >>>> particular step, by starting to set setting chooseleaf_vary_r=5 and then >>>> decrease it slowly to 1?) >>>> - then from firefly to hammer >>>> - then from hammer to jewel >>>> >>>> 2) or going directly to jewel tunables? >>>> >>>> Any advise on how to minimize the data movement? >>> >>> If you're switching tunables, there's going to be a ton of movement >>> and you need to prepare for it. But unlike with reweighting, there >>> isn't really a way to do it incrementally. I don't know if we have >>> experimental evidence but stepping through them is very unlikely to >>> help; I think you just want to pick the least bad time for a bunch of >>> migration and enable jewel tunables directly. >>> -Greg >> >> >> -- >> Goncalo Borges >> Research Computing >> ARC Centre of Excellence for Particle Physics at the Terascale >> School of Physics A28 | University of Sydney, NSW 2006 >> T: +61 2 93511937 >> > _______________________________________________ > ceph-users mailing list > ceph-users@xxxxxxxxxxxxxx > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |