Thanks Dan, After talking it through we've decided to adopt your approach too and leave the tunables till after the upgrade.
Regards, Adrian. On Mon, May 14, 2018 at 5:14 PM, Dan van der Ster <d...@vanderster.com> wrote: > Hi Adrian, > > Is there a strict reason why you *must* upgrade the tunables? > > It is normally OK to run with old (e.g. hammer) tunables on a luminous > cluster. The crush placement won't be state of the art, but that's not > a huge problem. > > We have a lot of data in a jewel cluster with hammer tunables. We'll > upgrade that to luminous soon, but don't plan to set chooseleaf_stable > until there's less disruptive procedure, e.g. [1]. > > Cheers, Dan > > [1] One idea I had to make this much less disruptive would be to > script something that uses upmap's to lock all PGs into their current > placement, then set chooseleaf_stable, then gradually remove the > upmap's. There are some details to work out, and it requires all > clients to be running luminous, but I think something like this could > help... > > > > > On Mon, May 14, 2018 at 9:01 AM, Adrian <aussie...@gmail.com> wrote: > > Hi all, > > > > We recently upgraded our old ceph cluster to jewel (5xmon, 21xstorage > hosts > > with 9x6tb filestore osds and 3xssd's with 3 journals on each) - mostly > used > > for openstack compute/cinder. > > > > In order to get there we had to go with chooseleaf_vary_r = 4 in order to > > minimize client impact and save time. We now need to get to luminous (on > a > > deadline and time is limited). > > > > Current tunables are: > > { > > "choose_local_tries": 0, > > "choose_local_fallback_tries": 0, > > "choose_total_tries": 50, > > "chooseleaf_descend_once": 1, > > "chooseleaf_vary_r": 4, > > "chooseleaf_stable": 0, > > "straw_calc_version": 1, > > "allowed_bucket_algs": 22, > > "profile": "unknown", > > "optimal_tunables": 0, > > "legacy_tunables": 0, > > "minimum_required_version": "firefly", > > "require_feature_tunables": 1, > > "require_feature_tunables2": 1, > > "has_v2_rules": 0, > > "require_feature_tunables3": 1, > > "has_v3_rules": 0, > > "has_v4_buckets": 0, > > "require_feature_tunables5": 0, > > "has_v5_rules": 0 > > } > > > > Setting chooseleaf_stable to 1, the crush compare tool says: > > Replacing the crushmap specified with --origin with the crushmap > > specified with --destination will move 8774 PGs (59.08417508417509% of > the > > total) > > from one item to another. > > > > Current tunings we have in ceph.conf are: > > #THROTTLING CEPH > > osd_max_backfills = 1 > > osd_recovery_max_active = 1 > > osd_recovery_op_priority = 1 > > osd_client_op_priority = 63 > > > > #PERFORMANCE TUNING > > osd_op_threads = 6 > > filestore_op_threads = 10 > > filestore_max_sync_interval = 30 > > > > I was wondering if anyone has any advice as to anything else we can do > > balancing client impact and speed of recovery or war stories of other > things > > to consider. > > > > I'm also wondering about the interplay between chooseleaf_vary_r and > > chooseleaf_stable. > > Are we better with > > 1) sticking with choosleaf_vary_r = 4, setting chooseleaf_stable =1, > > upgrading and then setting chooseleaf_vary_r incrementally to 1 when more > > time is available > > or > > 2) setting chooseleaf_vary_r incrementally first, then chooseleaf_stable > and > > finally upgrade > > > > All this bearing in mind we'd like to keep the time it takes us to get to > > luminous as short as possible ;-) (guestimating a 59% rebalance to take > many > > days) > > > > Any advice/thoughts gratefully received. > > > > Regards, > > Adrian. > > > > -- > > --- > > Adrian : aussie...@gmail.com > > If violence doesn't solve your problem, you're not using enough of it. > > > > _______________________________________________ > > ceph-users mailing list > > ceph-users@lists.ceph.com > > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > > > -- --- Adrian : aussie...@gmail.com If violence doesn't solve your problem, you're not using enough of it.
_______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com