[Gluster-users] Performance problems

[Vikas Gorur]

On Mar 29, 2010, at 9:48 AM, Olivier Le Cam wrote:
> gl2:/web-nunki# dd if=/dev/zero of=./test count=10000 bs=160000
> 10000+0 enregistrements lus
> 10000+0 enregistrements écrits
> 1600000000 bytes (1,6 GB) copied, 15,0147 s, 107 MB/s
> 
> gl2:/web-nunki# dd if=/dev/zero of=./test count=10000 bs=320000
> 10000+0 enregistrements lus
> 10000+0 enregistrements écrits
> 3200000000 bytes (3,2 GB) copied, 28,8645 s, 111 MB/s
> 
> TBH, I don't know what to conclude from these results.

Write performance increases with increasing block size because each individual write call
results in a network call which adds latency. Also, small block sizes mean more context switches
between kernel and GlusterFS. Higher block sizes result in fewer calls per byte
of data and thus give higher throughput. However, as you can see from your tests above,
the upward curve flattens out at 128k. This is because 128k is the largest chunk that the kernel
will send to GlusterFS. If you try with higher block sizes, the kernel breaks them up into 128k chunks
and thus you will see no more improvement.

------------------------------
Vikas Gorur
Engineer - Gluster, Inc.
+1 (408) 770 1894
------------------------------