You should do your reference test with dd with oflag=direct,dsync
direct will only bypass the cache while dsync will fsync on every
block which is much closer to reality of what ceph is doing afaik
On Thu, Jan 4, 2018 at 9:54 PM, Rafał Wądołowski
<rwadolowski@xxxxxxxxxxxxxx> wrote:
> Hi folks,
>
> I am currently benchmarking my cluster for an performance issue and I have
> no idea, what is going on. I am using these devices in qemu.
>
> Ceph version 12.2.2
>
> Infrastructure:
>
> 3 x Ceph-mon
>
> 11 x Ceph-osd
>
> Ceph-osd has 22x1TB Samsung SSD 850 EVO 1TB
>
> 96GB RAM
>
> 2x E5-2650 v4
>
> 4x10G Network (2 seperate bounds for cluster and public) with MTU 9000
>
>
> I had tested it with rados bench:
>
> # rados bench -p rbdbench 30 write -t 1
>
> Total time run: 30.055677
> Total writes made: 1199
> Write size: 4194304
> Object size: 4194304
> Bandwidth (MB/sec): 159.571
> Stddev Bandwidth: 6.83601
> Max bandwidth (MB/sec): 168
> Min bandwidth (MB/sec): 140
> Average IOPS: 39
> Stddev IOPS: 1
> Max IOPS: 42
> Min IOPS: 35
> Average Latency(s): 0.0250656
> Stddev Latency(s): 0.00321545
> Max latency(s): 0.0471699
> Min latency(s): 0.0206325
>
> # ceph tell osd.0 bench
> {
> "bytes_written": 1073741824,
> "blocksize": 4194304,
> "bytes_per_sec": 414199397
> }
>
> Testing osd directly
>
> # dd if=/dev/zero of=/dev/sdc bs=4M oflag=direct count=100
> 100+0 records in
> 100+0 records out
> 419430400 bytes (419 MB, 400 MiB) copied, 1.0066 s, 417 MB/s
>
> When I do dd inside vm (bs=4M wih direct), I have result like in rados
> bench.
>
> I think that the speed should be arround ~400MB/s.
>
> Is there any new parameters for rbd in luminous? Maybe I forgot about some
> performance tricks? If more information needed feel free to ask.
>
> --
> BR,
> Rafal Wadolowski
> _______________________________________________
> ceph-users mailing list
> ceph-users@xxxxxxxxxxxxxx
> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
