[Gluster-devel] Infiniband slow throughput

[nathan at robotics.net]

On Thu, 11 Sep 2008, Anand Avati wrote:

>> [root at cluster0 ~]# sync ; time (dd if=/dev/zero of=/share/foo.bar bs=4k
>> count=512K ; sync )
>> 524288+0 records in
>> 524288+0 records out
>> 2147483648 bytes (2.1 GB) copied, 158.164 seconds, 13.6 MB/s
>>
>> real    2m39.045s
>> user    0m0.108s
>> sys     0m1.420s
>
> write speeds without write-behind will be low for such small block sizes
> because of the high context switch overhead incurred by fuse. A bigger block
> size would help too. Please let us know the results with write-behind.

Writebehind showed almost no improvement at 4k block size, however moving 
from 4k to 128k blocks made a major difference, jumping to 136 MB/s. 
Anything else in my config that can speed things up? Should I be doing 
anything more client side?

http://share.robotics.net/client.vol
http://share.robotics.net/server.vol

><>
Nathan Stratton                                CTO, BlinkMind, Inc.
nathan at robotics.net                         nathan at blinkmind.com
http://www.robotics.net                        http://www.blinkmind.com