> I'm thinking of whether it makes sense to use the available SSDs in the > cluster nodes (1 SSD for 4 HDDs) as part of a writeback cache pool in front of > the IO intensive pool, instead of using them as journal SSDs? With this > method, the OSD journals would be co-located on the HDDs or the SSD:HDD > ratio could be reduced from 1:4 to something like 1:10. > > The write operations would still hit SSDs first (though latency would increase > compared to writing to dedicated SSD partitions local to the server), and as > far as I understand the cache flush operations are happening in a coalesced > fashion. > > Plus a definite advantage would be that besides functioning as a 'write log' > (aka. journal), the SSDs would be serving as a read cache for hot data. > > What do you think? > Have you considered bcache? It's in the kernel since 3.10 I think. It would be interesting to see comparisons between no ssd, journal on ssd, and bcache with ssd (with journal on same fs as osd) James
 |