On Thu, 27 Aug 2015, wenjunh wrote: > Hi > I have a try of the newstore to explore its performance, which shows > its performance is much poorer than filestore. > > I test my cluster using fio, Here is the comparison of the two store > with randread & randwrite scenario: > > rw=randread, bs=4K, numjobs=1: > newstore: bw=2280.7KB/s, iops=570 > filestore: bw=4846.6KB/s, iops=1211 > > rw=randwrite, bs=4K, numjobs=1: > newstore: bw=32999 B/s, iops=8 > filestore: bw=250978 B/s, iops=61 > The two tests are performed using the same hardware, but newstore is > much poorer than filestore. From the view of the community, newstore is > the next default backend store in Ceph, but why its performance so poor. > Could someone tell me why? Newstore doesn't do well with random IO just yet, especially when the WAL/journal is not on a separate device, because rocksdb doesn't do a very good job with its log management. I'm working on fixing this in rocksdb now. At the other end of the spectrum, if you do something like rados bench (4M writes) newstore should be almost 2x faster. Were you using an SSD journal with filestore or a disk partition? Newstore doesn't know how to use a separate SSD yet unless you set up the wal_dir manually. sage -- 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
