[Gluster-devel] Poor performance of block-store with RDMA
potatogim at gluesys.com
Thu Nov 9 10:29:39 UTC 2017
First of all, I really appreciate your test results for
My teammate and I tested block-store(glfs backstore with tcmu-runner)
but we have met a problem of performance.
We tested some cases with one server that has RDMA volume and one client
that is connected to same RDMA network.
two machines have same environment like below.
- Distro : CentOS 6.9
- Kernel : 4.12.9
- GlusterFS : 3.10.5
- tcmu-runner : 1.2.0
- iscsi-initiator-utils : 188.8.131.523
and these are results from test.
1st. FILEIO on FUSE mounted - 333MB/sec
2nd. glfs user backstore - 140MB/sec
3rd. FILEIO on FUSE mounted with tgtd - 235MB/sec
4th. glfs user backstore with tgtd - 220MB/sec
5th. FILEIO on FUSE mounted (iSER) - 643MB/sec
6th. glfs user backstore (iSER) - 149MB/sec
7th. FILEIO on FUSE mounted with tgtd (iSER) - 677MB/sec
8th. glfs user backstore with tgtd(iSER) - 535MB/sec
Every tests were done with dd command.
As shown above, 6th test showed similar results in 2nd test.
IMOH, It may be caused by tcmu-runner itself or glfs backstore handler
so I will take similar tests with other handlers(like qcow) in order to
find this point.
If there's anything I missed, Could you give me some tips for resolving
this issue? :-)
Research Engineer, Gluesys
Address. Gluesys R&D Center, 5F, 11-31, Simin-daero 327beon-gil,
E-Mail. potatogim at potatogim.net
The time I wasted today is the tomorrow the dead man was eager to see
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 516 bytes
Desc: OpenPGP digital signature
More information about the Gluster-devel