> On 26 Nov 2014, at 13:47, Christian Balzer <chibi@xxxxxxx> wrote: > > On Wed, 26 Nov 2014 05:37:43 -0600 Mark Nelson wrote: > >> On 11/26/2014 04:05 AM, Yujian Peng wrote: > [snip] >> >>> >>> Since the size of jornal partitions on SSDs is 10G, I want to set >>> filestore max sync interval to 30 minutes. Is 30 minutes resonable? >>> How to set filestore max sync interval on the fly? >> >> The effect of this is that you'll be able to write the journal >> uninterrupted for a long time, but syncs could be very expensive. I'd >> suggest giving it a try at a couple of values (default, 60s, 300s) under >> say 20 minutes of continuous heavy workload to observe the effects and >> see how you like it. >> > > When I asked more or less the same question a while ago, I was left with > the following information, please correct me if I'm wrong: > > 1. "filestore max sync interval" is the longest time one can write to the > journal, but flushes to the backing store can and will happen a lot > earlier than that. What triggers these besides what I'm going to list > below was left rather vague, though. If you want to enforce long periods > of journal usage before writing out and hopefully coalescing things a lot, > one needs to raise "filestore min sync interval". And the warnings about > expensive syncs go double when doing that. For any sync interval greater than ~30s you should also check and increase the kernel vm dirty buffer flushing period. See: vm.dirty_background_ratio vm.dirty_background_bytes vm.dirty_ratio vm.dirty_bytes vm.dirty_writeback_centisecs vm.dirty_expire_centisecs Cheers, Dan > 2. Once the journal is half full, a flush is triggered in any case. > > And once the journals are on SSDs, Yujian will likely discover that now > his CPUs are the bottleneck in certain situations. ^o^ > > As for setting things on the fly, look for the "injectargs" syntax, like: > ceph tell osd.* injectargs '--filestore_max_sync_interval 30' > > > Christian > -- > Christian Balzer Network/Systems Engineer > chibi@xxxxxxx Global OnLine Japan/Fusion Communications > http://www.gol.com/ > _______________________________________________ > 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