On Mon, Nov 29, 2010 at 6:55 PM, Jeff Wu <cpwu@xxxxxxxxxxxxx> wrote: >> Could you run "ceph osd tell * bench", then run "ceph -w", and report >> the results? (That'll just run local benchmarking on the OSD to report >> the approximate write speed it's capable of.) > > I run six times for the command: "$ sudo ceph osd tell 0/1 bench " > use "$ sudo ceph -w" to get the following results: > > osd0 : [INF] bench: wrote 1024 MB in blocks of 4096 KB in 12.906598 sec > at 49775 KB/sec > osd1 : [INF] bench: wrote 1024 MB in blocks of 4096 KB in 21.023294 sec > at 49384 KB/sec > osd0 : [INF] bench: wrote 1024 MB in blocks of 4096 KB in 12.834682 sec > at 51535 KB/sec > osd1 : [INF] bench: wrote 1024 MB in blocks of 4096 KB in 20.792697 sec > at 37547 KB/sec > osd0 : [INF] bench: wrote 1024 MB in blocks of 4096 KB in 13.058412 sec > at 77191 KB/sec > osd1 : [INF] bench: wrote 1024 MB in blocks of 4096 KB in 21.113612 sec > at 47369 KB/sec Okay, those are a bit slow but reasonable. Based on these I'd expect you to generally manage about 40-50MB/s (since everything is replicated; in a 2-disk configuration it'll just be the speed of your slowest disk), assuming a properly configured system. >> You can also run "rados -p data bench 60 write", and then "rados -p >> data bench 0 seq" to get a simpler (better understood) performance >> test. > > I run twice for the command: "rados -p data bench 60 write" , > Get the results: > $ sudo rados -p data bench 60 write > .......................... > .......................... > > Total time run: 76.182225 > Total writes made: 121 > Write size: 4194304 > Bandwidth (MB/sec): 6.219 > > Average Latency: 13.3068 > Max latency: 23.9986 > Min latency: 7.01847 WOAH. That's a lot of latency. Rather more than I'd expect to get just from seek times in a non-journaled environment. What's the round-trip time to ping the OSDs from your client? Are your disks okay? > but run "$ sudo rados -p data bench 0 seq" > fail to get the results, Maybe it's a bug,ceph version 0.23. Oh right, sorry. We put in a quick fix to make the write benchmark scale across multiple client writers and forgot to adjust the read benchmark. *oops* -- 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
