Hello folks,
I am new to ceph and at the moment I am doing some performance tests with a 4 node
ceph-cluster (pacific, 16.2.1).
Node hardware (4 identical nodes):
* DELL 3620 workstation
* Intel Quad-Core i7-6700(a)3.4 GHz
* 8 GB RAM
* Debian Buster (base system, installed a dedicated on Patriot Burst 120 GB SATA-SSD)
* HP 530SPF+ 10 GBit dual-port NIC (tested with iperf to 9.4 GBit/s from node to
node)
* 1 x Kingston KC2500 M2 NVMe PCIe SSD (500 GB, NO power loss protection !)
* 3 x Seagate Barracuda SATA disk drives (7200 rpm, 500 GB)
After bootstrapping a containerized (docker) ceph-cluster, I did some performance tests on
the NVMe storage by creating a storage pool called „ssdpool“, consisting of 4 OSDs per
(one) NVMe device (per node). A first write-performance test yields
=============
root@ceph1:~# rados bench -p ssdpool 10 write -b 4M -t 16 --no-cleanup
hints = 1
Maintaining 16 concurrent writes of 4194304 bytes to objects of size 4194304 for up to 10
seconds or 0 objects
Object prefix: benchmark_data_ceph1_78
sec Cur ops started finished avg MB/s cur MB/s last lat(s) avg lat(s)
0 0 0 0 0 0 - 0
1 16 30 14 55.997 56 0.0209977 0.493427
2 16 53 37 73.9903 92 0.0264305 0.692179
3 16 76 60 79.9871 92 0.559505 0.664204
4 16 99 83 82.9879 92 0.609332 0.721016
5 16 116 100 79.9889 68 0.686093 0.698084
6 16 132 116 77.3224 64 1.19715 0.731808
7 16 153 137 78.2741 84 0.622646 0.755812
8 16 171 155 77.486 72 0.25409 0.764022
9 16 192 176 78.2076 84 0.968321 0.775292
10 16 214 198 79.1856 88 0.401339 0.766764
11 1 214 213 77.4408 60 0.969693 0.784002
Total time run: 11.0698
Total writes made: 214
Write size: 4194304
Object size: 4194304
Bandwidth (MB/sec): 77.3272
Stddev Bandwidth: 13.7722
Max bandwidth (MB/sec): 92
Min bandwidth (MB/sec): 56
Average IOPS: 19
Stddev IOPS: 3.44304
Max IOPS: 23
Min IOPS: 14
Average Latency(s): 0.785372
Stddev Latency(s): 0.49011
Max latency(s): 2.16532
Min latency(s): 0.0144995
=============
... and I think that 80 MB/s throughput is a very poor result in conjunction with NVMe
devices and 10 GBit nics.
A bare write-test (with fsync=0 option) of the NVMe drives yields a write throughput of
round about 800 MB/s per device ... the second test (with fsync=1) drops performance to
200 MB/s.
=============
root@ceph1:/home/mschmid# fio --rw=randwrite --name=IOPS-write --bs=1024k --direct=1
--filename=/dev/nvme0n1 --numjobs=4 --ioengine=libaio --iodepth=32 --refill_buffers
--group_reporting --runtime=30 --time_based --fsync=0
IOPS-write: (g=0): rw=randwrite, bs=(R) 1024KiB-1024KiB, (W) 1024KiB-1024KiB, (T)
1024KiB-1024KiB, ioengine=libaio, iodepth=32...
fio-3.12
Starting 4 processes
Jobs: 4 (f=4): [w(4)][100.0%][w=723MiB/s][w=722 IOPS][eta 00m:00s]
IOPS-write: (groupid=0, jobs=4): err= 0: pid=31585: Thu Apr 29 15:15:03 2021
write: IOPS=740, BW=740MiB/s (776MB/s)(21.8GiB/30206msec); 0 zone resets
slat (usec): min=16, max=810, avg=106.48, stdev=30.48
clat (msec): min=7, max=1110, avg=172.09, stdev=120.18
lat (msec): min=7, max=1110, avg=172.19, stdev=120.18
clat percentiles (msec):
| 1.00th=[ 32], 5.00th=[ 48], 10.00th=[ 53], 20.00th=[ 63],
| 30.00th=[ 115], 40.00th=[ 161], 50.00th=[ 169], 60.00th=[ 178],
| 70.00th=[ 190], 80.00th=[ 220], 90.00th=[ 264], 95.00th=[ 368],
| 99.00th=[ 667], 99.50th=[ 751], 99.90th=[ 894], 99.95th=[ 986],
| 99.99th=[ 1036]
bw ( KiB/s): min=22528, max=639744, per=25.02%, avg=189649.94, stdev=113845.69,
samples=240
iops : min= 22, max= 624, avg=185.11, stdev=111.18, samples=240
lat (msec) : 10=0.01%, 20=0.19%, 50=6.43%, 100=20.29%, 250=61.52%
lat (msec) : 500=8.21%, 750=2.85%, 1000=0.47%
cpu : usr=11.87%, sys=2.05%, ctx=13141, majf=0, minf=45
IO depths : 1=0.1%, 2=0.1%, 4=0.1%, 8=0.1%, 16=0.3%, 32=99.4%, >=64=0.0%
submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.1%, 64=0.0%, >=64=0.0%
issued rwts: total=0,22359,0,0 short=0,0,0,0 dropped=0,0,0,0
latency : target=0, window=0, percentile=100.00%, depth=32
Run status group 0 (all jobs):
WRITE: bw=740MiB/s (776MB/s), 740MiB/s-740MiB/s (776MB/s-776MB/s), io=21.8GiB (23.4GB),
run=30206-30206msec
Disk stats (read/write):
nvme0n1: ios=0/89150, merge=0/0, ticks=0/15065724, in_queue=15118720, util=99.75%
=============
Furthermore an IOPS-test on the NVMe device with block-size 4k shows round about 1000 IOPS
with fsnyc=1 and 35000 IOPS with fsync=0.
To my question: As CPU- and network-load seem to be low during my tests, I would like to
know, which bottleneck can cause such a huge performance drop between the bare
hardware-performance of the nvme-drives and the write-speeds in the rados benchmark. Could
the missing power loss protection (fsync=1) be the problem, or what throughput should one
expect to be normal in such a setup?
Thanks for every advice!
Best regards,
Michael