Hi! I can post the crush map tomorrow morning, but it definitely isn't targeting the NVME drives. I'm having a performance issue specifically with the HDD-backed pool, where each OSD is an NVME-backed WAL/DB + HDD-backed storage. /Z On Tue, 5 Oct 2021, 22:43 Tor Martin Ølberg, <tmolberg@xxxxxxxxx> wrote: > Hi Zakhar, > > Out of curiosity, what does your crushmap look like? Probably a long shot > but are you sure your crush map is targeting the NVME's for the rados bench > you are performing? > > Tor Martin Ølberg > > On Tue, Oct 5, 2021 at 9:31 PM Christian Wuerdig < > christian.wuerdig@xxxxxxxxx> wrote: > >> Maybe some info is missing but 7k write IOPs at 4k block size seem fairly >> decent (as you also state) - the bandwidth automatically follows from that >> so not sure what you're expecting? >> I am a bit puzzled though - by my math 7k IOPS at 4k should only be >> 27MiB/sec - not sure how the 120MiB/sec was achieved >> The read benchmark seems in line with 13k IOPS at 4k making around >> 52MiB/sec bandwidth which again is expected. >> >> >> On Wed, 6 Oct 2021 at 04:08, Zakhar Kirpichenko <zakhar@xxxxxxxxx> wrote: >> >> > Hi, >> > >> > I built a CEPH 16.2.x cluster with relatively fast and modern hardware, >> and >> > its performance is kind of disappointing. I would very much appreciate >> an >> > advice and/or pointers :-) >> > >> > The hardware is 3 x Supermicro SSG-6029P nodes, each equipped with: >> > >> > 2 x Intel(R) Xeon(R) Gold 5220R CPUs >> > 384 GB RAM >> > 2 x boot drives >> > 2 x 1.6 TB Micron 7300 MTFDHBE1T6TDG drives (DB/WAL) >> > 2 x 6.4 TB Micron 7300 MTFDHBE6T4TDG drives (storage tier) >> > 9 x Toshiba MG06SCA10TE 9TB HDDs, write cache off (storage tier) >> > 2 x Intel XL710 NICs connected to a pair of 40/100GE switches >> > >> > All 3 nodes are running Ubuntu 20.04 LTS with the latest 5.4 kernel, >> > apparmor is disabled, energy-saving features are disabled. The network >> > between the CEPH nodes is 40G, CEPH access network is 40G, the average >> > latencies are < 0.15 ms. I've personally tested the network for >> throughput, >> > latency and loss, and can tell that it's operating as expected and >> doesn't >> > exhibit any issues at idle or under load. >> > >> > The CEPH cluster is set up with 2 storage classes, NVME and HDD, with 2 >> > smaller NVME drives in each node used as DB/WAL and each HDD allocated . >> > ceph osd tree output: >> > >> > ID CLASS WEIGHT TYPE NAME STATUS REWEIGHT >> PRI-AFF >> > -1 288.37488 root default >> > -13 288.37488 datacenter ste >> > -14 288.37488 rack rack01 >> > -7 96.12495 host ceph01 >> > 0 hdd 9.38680 osd.0 up 1.00000 >> 1.00000 >> > 1 hdd 9.38680 osd.1 up 1.00000 >> 1.00000 >> > 2 hdd 9.38680 osd.2 up 1.00000 >> 1.00000 >> > 3 hdd 9.38680 osd.3 up 1.00000 >> 1.00000 >> > 4 hdd 9.38680 osd.4 up 1.00000 >> 1.00000 >> > 5 hdd 9.38680 osd.5 up 1.00000 >> 1.00000 >> > 6 hdd 9.38680 osd.6 up 1.00000 >> 1.00000 >> > 7 hdd 9.38680 osd.7 up 1.00000 >> 1.00000 >> > 8 hdd 9.38680 osd.8 up 1.00000 >> 1.00000 >> > 9 nvme 5.82190 osd.9 up 1.00000 >> 1.00000 >> > 10 nvme 5.82190 osd.10 up 1.00000 >> 1.00000 >> > -10 96.12495 host ceph02 >> > 11 hdd 9.38680 osd.11 up 1.00000 >> 1.00000 >> > 12 hdd 9.38680 osd.12 up 1.00000 >> 1.00000 >> > 13 hdd 9.38680 osd.13 up 1.00000 >> 1.00000 >> > 14 hdd 9.38680 osd.14 up 1.00000 >> 1.00000 >> > 15 hdd 9.38680 osd.15 up 1.00000 >> 1.00000 >> > 16 hdd 9.38680 osd.16 up 1.00000 >> 1.00000 >> > 17 hdd 9.38680 osd.17 up 1.00000 >> 1.00000 >> > 18 hdd 9.38680 osd.18 up 1.00000 >> 1.00000 >> > 19 hdd 9.38680 osd.19 up 1.00000 >> 1.00000 >> > 20 nvme 5.82190 osd.20 up 1.00000 >> 1.00000 >> > 21 nvme 5.82190 osd.21 up 1.00000 >> 1.00000 >> > -3 96.12495 host ceph03 >> > 22 hdd 9.38680 osd.22 up 1.00000 >> 1.00000 >> > 23 hdd 9.38680 osd.23 up 1.00000 >> 1.00000 >> > 24 hdd 9.38680 osd.24 up 1.00000 >> 1.00000 >> > 25 hdd 9.38680 osd.25 up 1.00000 >> 1.00000 >> > 26 hdd 9.38680 osd.26 up 1.00000 >> 1.00000 >> > 27 hdd 9.38680 osd.27 up 1.00000 >> 1.00000 >> > 28 hdd 9.38680 osd.28 up 1.00000 >> 1.00000 >> > 29 hdd 9.38680 osd.29 up 1.00000 >> 1.00000 >> > 30 hdd 9.38680 osd.30 up 1.00000 >> 1.00000 >> > 31 nvme 5.82190 osd.31 up 1.00000 >> 1.00000 >> > 32 nvme 5.82190 osd.32 up 1.00000 >> 1.00000 >> > >> > ceph df: >> > >> > --- RAW STORAGE --- >> > CLASS SIZE AVAIL USED RAW USED %RAW USED >> > hdd 253 TiB 241 TiB 13 TiB 13 TiB 5.00 >> > nvme 35 TiB 35 TiB 82 GiB 82 GiB 0.23 >> > TOTAL 288 TiB 276 TiB 13 TiB 13 TiB 4.42 >> > >> > --- POOLS --- >> > POOL ID PGS STORED OBJECTS USED %USED MAX >> AVAIL >> > images 12 256 24 GiB 3.15k 73 GiB 0.03 76 >> TiB >> > volumes 13 256 839 GiB 232.16k 2.5 TiB 1.07 76 >> TiB >> > backups 14 256 31 GiB 8.56k 94 GiB 0.04 76 >> TiB >> > vms 15 256 752 GiB 198.80k 2.2 TiB 0.96 76 >> TiB >> > device_health_metrics 16 32 35 MiB 39 106 MiB 0 76 >> TiB >> > volumes-nvme 17 256 28 GiB 7.21k 81 GiB 0.24 11 >> TiB >> > ec-volumes-meta 18 256 27 KiB 4 92 KiB 0 76 >> TiB >> > ec-volumes-data 19 256 8 KiB 1 12 KiB 0 152 >> TiB >> > >> > Please disregard the ec-pools, as they're not currently in use. All >> other >> > pools are configured with min_size=2, size=3. All pools are bound to HDD >> > storage except for 'volumes-nvme', which is bound to NVME. The number of >> > PGs was increased recently, as with autoscaler I was getting a very >> uneven >> > PG distribution on devices and we're expecting to add 3 more nodes of >> > exactly the same configuration in the coming weeks. I have to emphasize >> > that I tested different PG numbers and they didn't have a noticeable >> impact >> > on the cluster performance. >> > >> > The main issue is that this beautiful cluster isn't very fast. When I >> test >> > against the 'volumes' pool, residing on HDD storage class (HDDs with >> DB/WAL >> > on NVME), I get unexpectedly low throughput numbers: >> > >> > > rados -p volumes bench 30 write --no-cleanup >> > ... >> > Total time run: 30.3078 >> > Total writes made: 3731 >> > Write size: 4194304 >> > Object size: 4194304 >> > Bandwidth (MB/sec): 492.415 >> > Stddev Bandwidth: 161.777 >> > Max bandwidth (MB/sec): 820 >> > Min bandwidth (MB/sec): 204 >> > Average IOPS: 123 >> > Stddev IOPS: 40.4442 >> > Max IOPS: 205 >> > Min IOPS: 51 >> > Average Latency(s): 0.129115 >> > Stddev Latency(s): 0.143881 >> > Max latency(s): 1.35669 >> > Min latency(s): 0.0228179 >> > >> > > rados -p volumes bench 30 seq --no-cleanup >> > ... >> > Total time run: 14.7272 >> > Total reads made: 3731 >> > Read size: 4194304 >> > Object size: 4194304 >> > Bandwidth (MB/sec): 1013.36 >> > Average IOPS: 253 >> > Stddev IOPS: 63.8709 >> > Max IOPS: 323 >> > Min IOPS: 91 >> > Average Latency(s): 0.0625202 >> > Max latency(s): 0.551629 >> > Min latency(s): 0.010683 >> > >> > On average, I get around 550 MB/s writes and 800 MB/s reads with 16 >> threads >> > and 4MB blocks. The numbers don't look fantastic for this hardware, I >> can >> > actually push over 8 GB/s of throughput with fio, 16 threads and 4MB >> blocks >> > from an RBD client (KVM Linux VM) connected over a low-latency 40G >> network, >> > probably hitting some OSD caches there: >> > >> > READ: bw=8525MiB/s (8939MB/s), 58.8MiB/s-1009MiB/s >> (61.7MB/s-1058MB/s), >> > io=501GiB (538GB), run=60001-60153msec >> > Disk stats (read/write): >> > vdc: ios=48163/0, merge=6027/0, ticks=1400509/0, in_queue=1305092, >> > util=99.48% >> > >> > The issue manifests when the same client does something closer to >> real-life >> > usage, like a single-thread write or read with 4KB blocks, as if using >> for >> > example ext4 file system: >> > >> > > fio --name=ttt --ioengine=posixaio --rw=write --bs=4k --numjobs=1 >> > --size=4g --iodepth=1 --runtime=60 --time_based --end_fsync=1 >> > ... >> > Run status group 0 (all jobs): >> > WRITE: bw=120MiB/s (126MB/s), 120MiB/s-120MiB/s (126MB/s-126MB/s), >> > io=7694MiB (8067MB), run=64079-64079msec >> > Disk stats (read/write): >> > vdc: ios=0/6985, merge=0/406, ticks=0/3062535, in_queue=3048216, >> > util=77.31% >> > >> > > fio --name=ttt --ioengine=posixaio --rw=read --bs=4k --numjobs=1 >> > --size=4g --iodepth=1 --runtime=60 --time_based --end_fsync=1 >> > ... >> > Run status group 0 (all jobs): >> > READ: bw=54.0MiB/s (56.7MB/s), 54.0MiB/s-54.0MiB/s >> (56.7MB/s-56.7MB/s), >> > io=3242MiB (3399MB), run=60001-60001msec >> > Disk stats (read/write): >> > vdc: ios=12952/3, merge=0/1, ticks=81706/1, in_queue=56336, >> util=99.13% >> > >> > And this is a total disaster: the IOPS look decent, but the bandwidth is >> > unexpectedly very very low. I just don't understand why a single RBD >> client >> > writes at 120 MB/s (sometimes slower), and 50 MB/s reads look like a bad >> > joke ¯\_(ツ)_/¯ >> > >> > When I run these benchmarks, nothing seems to be overloaded, things like >> > CPU and network are barely utilized, OSD latencies don't show anything >> > unusual. Thus I am puzzled with these results, as in my opinion SAS HDDs >> > with DB/WAL on NVME drives should produce better I/O bandwidth, both for >> > writes and reads. I mean, I can easily get much better performance from >> a >> > single HDD shared over network via NFS or iSCSI. >> > >> > I am open to suggestions and would very much appreciate comments and/or >> an >> > advice on how to improve the cluster performance. >> > >> > Best regards, >> > Zakhar >> > _______________________________________________ >> > ceph-users mailing list -- ceph-users@xxxxxxx >> > To unsubscribe send an email to ceph-users-leave@xxxxxxx >> > >> _______________________________________________ >> ceph-users mailing list -- ceph-users@xxxxxxx >> To unsubscribe send an email to ceph-users-leave@xxxxxxx >> > _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx
 |