[Gluster-users] glsuterfs cpu parallelism option?

Tobias Winter tobias at linuxdingsda.de
Thu Mar 14 12:10:22 UTC 2013


Hi,

it seems the problem only occurs when dealing with a lot of small files.
The performance drops to ~30-50MB/s while the glusterfs tasks peaks at
200%CPU on the brick where the data is being written to gluster.

Writing big files still isn't nearly as fast as the blockdevice but at
least it can easily handle 450-500mb/s which is about 50% of hardware
throughput capability.

We fiddled a bit with gluster and were hoping that we could reduce CPU
load but so far with no luck. Does anyone have some insight into how to
reduce the cpu load of glusterfs or how to spread it over more than 2 CPUS?

Is that something that gets better with the current beta release?

backup-eu1:~# gluster volume info gv0

Volume Name: gv0
Type: Replicate
Volume ID: 0d2e9442-8109-46db-bd6a-84d4f2148bb3
Status: Started
Number of Bricks: 1 x 2 = 2
Transport-type: tcp
Bricks:
Brick1: backup-eu1.lan:/raid0/glusterstore
Brick2: backup-eu2.lan:/raid0/glusterstore
Options Reconfigured:
performance.cache-size: 256MB
backup-eu1:~#

global config file options:

    option transport-type tcp
    option transport.socket.keepalive-time 10
    option transport.socket.nodelay on
    option transport.socket.keepalive-interval 2
    option transport.socket.read-fail-log off

Any hints are greatly appreciated.

Thanks

Tobias



On 03/13/2013 02:07 PM, Tobias Winter wrote:
> Hi,
> 
> we're setting up a glusterfs server apir with replication and am testing
> performance. While doing that we stumbled over that in top:
> 
>   PID USER      PR  NI  VIRT  RES  SHR S  %CPU %MEM    TIME+  COMMAND
> 
>  5151 root      20   0 6184m 5.9g 2540 R   200  9.3  28:27.99 glusterfs
> 
>  5618 root      20   0  392m  34m 2116 S    38  0.1  24:53.10 glusterfsd
> 
> 
> When under heavy load, CPU usage will not exceed 200% hence I'm assuming
> it only spawns two threads. Can that be tweaked?
> 
> My hope is that when that is increased and glusterfs can use more cpus,
> the overall performance will increase drastically.
> 
> Thanks
> 
> Tobias
> 
> 
> 
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://supercolony.gluster.org/mailman/listinfo/gluster-users
> 

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 263 bytes
Desc: OpenPGP digital signature
URL: <http://supercolony.gluster.org/pipermail/gluster-users/attachments/20130314/459946a1/attachment.sig>


More information about the Gluster-users mailing list