[Gluster-users] Really slow performance

[Harald Stürzebecher]

Hello!

2008/12/12 Matt Lawrence <matt.lawrence at tamu.edu>:
> I am seeing extremely slow performance with glusterfs.

[...]

> Writing directly to the underlying filesystem:
>
> -bash-3.2# dd if=/dev/zero of=/glusterfs/shelf-1/testing count=20M
[...]
> 10737418240 bytes (11 GB) copied, 230.862 seconds, 46.5 MB/s
>
>
> Writing through glusterfs with no readahead/writebehind:
>
> -bash-3.2# dd if=/dev/zero of=/storage/testing count=20M
[...]
> 10737418240 bytes (11 GB) copied, 3429.89 seconds, 3.1 MB/s
>
>
> Using the above configuration is even slower:
>
> -bash-3.2# dd if=/dev/zero of=/storage/testing count=20M
[...]
> 10737418240 bytes (11 GB) copied, 4379.16 seconds, 2.5 MB/s
>
>
> What should I be trying to improve performance?  I really need help!

IIRC:
- GlusterFS 1.3.x has problems with communication overhead. As a
result, transporting small blocks of data is very slow.
- (11GB data /20M blocks ~0,5k blocksize) looks small to me

I'd try a bigger blocksize with dd to check if that really is the
problem. (e.g. -bs=64k or even -bs=1M)
If that helps, I'd try writebehind on the client to let GlusterFS
transport bigger blocks.

http://www.gluster.org/docs/index.php/GlusterFS_Translators_v1.3#Write_Behind_Translator


Harald Stürzebecher