We do observe the same issue on our 12 SSD setup, disable the all log maybe helpful.
Cheers, xinxin From: ceph-users [mailto:ceph-users-boun...@lists.ceph.com] On Behalf Of Mark Wu Sent: Friday, October 17, 2014 12:18 AM To: ceph-users@lists.ceph.com Subject: [ceph-users] Performance doesn't scale well on a full ssd cluster. Hi list, During my test, I found ceph doesn't scale as I expected on a 30 osds cluster. The following is the information of my setup: HW configuration: 15 Dell R720 servers, and each server has: Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz, 20 cores and hyper-thread enabled. 128GB memory two Intel 3500 SSD disks, connected with MegaRAID SAS 2208 controller, each disk is configured as raid0 separately. bonding with two 10GbE nics, used for both the public network and cluster network. SW configuration: OS CentOS 6.5, Kernel 3.17, Ceph 0.86 XFS as file system for data. each SSD disk has two partitions, one is osd data and the other is osd journal. the pool has 2048 pgs. 2 replicas. 5 monitors running on 5 of the 15 servers. Ceph configuration (in memory debugging options are disabled) [osd] osd data = /var/lib/ceph/osd/$cluster-$id osd journal = /var/lib/ceph/osd/$cluster-$id/journal osd mkfs type = xfs osd mkfs options xfs = -f -i size=2048 osd mount options xfs = rw,noatime,logbsize=256k,delaylog osd journal size = 20480 osd mon heartbeat interval = 30 # Performance tuning filestore osd_max_backfills = 10 osd_recovery_max_active = 15 merge threshold = 40 filestore split multiple = 8 filestore fd cache size = 1024 osd op threads = 64 # Recovery tuning osd recovery max active = 1 osd max backfills = 1 osd recovery op priority = 1 throttler perf counter = false osd enable op tracker = false filestore_queue_max_ops = 5000 filestore_queue_committing_max_ops = 5000 journal_max_write_entries = 1000 journal_queue_max_ops = 5000 objecter_inflight_ops = 8192 When I test with 7 servers (14 osds), the maximum iops of 4k random write I saw is 17k on single volume and 44k on the whole cluster. I expected the number of 30 osds cluster could approximate 90k. But unfornately, I found that with 30 osds, it almost provides the performce as 14 osds, even worse sometime. I checked the iostat output on all the nodes, which have similar numbers. It's well distributed but disk utilization is low. In the test with 14 osds, I can see higher utilization of disk (80%~90%). So do you have any tunning suggestion to improve the performace with 30 osds? Any feedback is appreciated. iostat output: Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 sdb 0.00 88.50 0.00 5188.00 0.00 93397.00 18.00 0.90 0.17 0.09 47.85 sdc 0.00 443.50 0.00 5561.50 0.00 97324.00 17.50 4.06 0.73 0.09 47.90 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.00 17.50 0.00 28.00 0.00 3948.00 141.00 0.01 0.29 0.05 0.15 sdb 0.00 69.50 0.00 4932.00 0.00 87067.50 17.65 2.27 0.46 0.09 43.45 sdc 0.00 69.00 0.00 4855.50 0.00 105771.50 21.78 0.95 0.20 0.10 46.40 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 dm-1 0.00 0.00 0.00 42.50 0.00 3948.00 92.89 0.01 0.19 0.04 0.15 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.00 12.00 0.00 8.00 0.00 568.00 71.00 0.00 0.12 0.12 0.10 sdb 0.00 72.50 0.00 5046.50 0.00 113198.50 22.43 1.09 0.22 0.10 51.40 sdc 0.00 72.50 0.00 4912.00 0.00 91204.50 18.57 2.25 0.46 0.09 43.60 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 dm-1 0.00 0.00 0.00 18.00 0.00 568.00 31.56 0.00 0.17 0.06 0.10 Regards, Mark Wu
_______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com