sorry, i just realized the deferrd ios in a batch already be ordered in DeferredBatch::prepare_write. i saw the seastar impl a userspace disk io scheduler with multiple io queues, if we provide each pg sharded thread with one io queue, can we gain some benefit on performance? i think the userspace scheduler's priority class also could be used by ceph. .. 2018-03-30 17:43 GMT+08:00 zengran zhang <z13121369189@xxxxxxxxx>: > Hi Sage: > now the direct write and the deferred write is both submit through > libaio with O_DIRECT, is it possible to submit the deferred buffered. > the fsync of both direct and deferred is in the sync_thread. maybe we > could call the fadvance after the fsync with POSIX_FADV_DONTNEED to > release the deferred ios's pages. but i dont kown wether is it safe > when reading? i kown we need be careful when we mix the buffered io > and the direct io. > > the benefit of doing this may huge on hdd environment because the > kernel block scheduler may merge small ios or order them... -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
