On Sat 2015-Nov-07 09:24:06 +0100, Rémi BUISSON <remi-buisson@xxxxxxxxx> wrote:
Hi guys, I would need your help to figure out performance issues on my ceph cluster.I've read pretty much every thread on the net concerning this topic but I didn't manage to have acceptable performances. In my company, we are planning to replace our existing virtualization infrastucture NAS by a ceph cluster in order to improve the global platform performances, scalability and security. The current NAS we have handle about 50k iops.For this we bought:2 x NFS servers: 2 x Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz, 32 GB RAM, 2 x 10Gbps network interfaces (bonding) 3 x MON servers: 1 x Intel(R) Xeon(R) CPU E5-2603 v3 @ 1.60GHz, 16 GB RAM, 2 x 10Gbps network interfaces (bonding) 2 x MDS servers: 2 x Intel(R) Xeon(R) CPU E5-2687W v3 @ 3.10GHz, 32 GB RAM, 2 x 10Gbps network interfaces (bonding) 2 x OSD servers (cache): 2 x Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz, 256 GB RAM, 2 x SSD INTEL SSDSC2BX200G4 (200 GB) for journal, 6 x SSD INTEL SSDSC2BX016T4R (1,4 TB) for data, 2 x 10Gbps network interfaces (bonding) 4 x OSD servers (storage): 2 x Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz, 256 GB RAM, 4 x SSD TOSHIBA PX02SMF020 (200GB) for journal, 18 x HGST Ultrastar HUC101818CS4204 (1.8TB) for data, 2 x 10Gbps network interfaces (bonding)The total of this is 84 OSDs.I created two 4096 pgs pools, one called rbd-cold-storage and the other rbd-hot-storage. As you may guess, the rbd-cold-storage is composed of the 4 OSD servers with platter disks and the rbd-hot-storage is composed of the 2 OSD servers with SSD disks. On the rdb-cold-storage, I created an rbd device which is mapped on the NFS server.I benched each of the SSD we have and it can handle 40k iops each. As my replication factor is 2, the theoritical performance of the cluster is (2 x 6 (OSD cache) x 40k) / 2 = 240k iops.
Aside from the other more detailed replies re: tuning, isn't the layout of the caching tier journals sub-optimal in this scenario? Given the similar model numbers there, I'm assuming the performance (throughput, IOPS) of the journal & data disks are similar, but please correct me if I'm wrong there. My understanding of ceph's design (newer to ceph; please excuse misunderstandings) is that writes pass through the journals, the OSD will ACK writes when they are committed to the journal(s) of the OSDs in that PG (so, one other OSD in this case with a replication factor of 2), and journals are then flushed to OSDs asynchronously.
Rather than "(2 x 6 (OSD cache) x 40k) / 2 = 240k iops", isn't the calculation actually:
(# hosts) x (# journal disks) x (IOPS per journal disk) / (replication factor) ? IOW: (2 x 2 (OSD cache journal SSDs) x 40K) / 2 = 80KYes, putting journals on the same disk as the OSD's data halves your write performance because data has to flush from the journal partition to the data partition on the same SSD, but in this case would it not be more optimal to just chuck the 2x SSDSC2BX200G4 per cache host, replace them with 2x more data disks (SSDSC2BX016T4R) for 8 total per cache OSD host, and then go with journals on the same disk?
In that case we're looking at: (# hosts) x (# journal disks) x (IOPS per journal disk) / (replication factor) / 2...where the final division by 2 is our write penalty for sharing journal and data on the same disk.
So, in this scenario: 2 x 8 (OSD cache SSDs) x 40K / 2 (replication factor) / 2 = 160K Yes/no?In a regular "SSD journals + spinners for data" setup, journals on discrete/partitioned SSDs makes sense in e.g. a 3:1 ratio as your performance (well, throughput; IOPS is another story) on the SSD will generally be ~3x what your SAS/SATA spinners can do. So: 1 SSD has 3x partitions and serves journals for 3x OSDs backed by spinners, the numbers are matched up so that it has the capacity to absorb (write) data as fast as it can flush it down to the spinners and it can pretty much max the spinnners' write capacity. Overload the SSD with too many journals and it will be maxed with spinners sitting waiting/idle.
But in scenarios where the performance of the journal SSDs matches the performance of the backing disks for the OSDs and with a 3:1 ratio on data disks to journal disks, the data SSDs will still have more write performance capacity to spare while the journal SSD is maxed. Don't we need something with greater throughput/IOPS in the journal than in our data partition in order to make discrete journals be of benefit?
I guess the alternative to swapping the 2x SSDSC2BX200G4 journals in the cache for simply more data disks (SSDSC2BX016T4R) would be to go PCIe/NVMe for the journals in the cache layer, at which point the discrete journals could be a net plus again?
-- Hugo cell: 604-617-3133 hugo@xxxxxxxxxxx: email, xmpp/jabber PGP fingerprint (B178313E): CF18 15FA 9FE4 0CD1 2319 1D77 9AB1 0FFD B178 313E (also on Signal)
Attachment:
signature.asc
Description: Digital signature
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
