[Gluster-users] Why so bad performance?

[Rash]

sob, 06 gru 2008, 12:51:08 +0800, Kirby Zhou napisał(a):
> I have tried.
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> volume writebehind
>   type performance/write-behind
>   option aggregate-size 8MB # default is 0bytes
>   option flush-behind off   # default is 'off'
>   subvolumes unify0
> end-volume
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> [@123.25 ~]# dd if=/opt/xxx of=/mnt/xxx
> 100353+0 records in
> 100353+0 records out
> 51380736 bytes (51 MB) copied, 129.343 seconds, 397 kB/s
> 
> [@123.25 ~]# dd if=/opt/xxx of=/mnt/xxx bs=16M
> 16+0 records in
> 16+0 records out
> 268435456 bytes (268 MB) copied, 23.5073 seconds, 11.4 MB/s

I've got almost the same problem on 1950 dell servers connected with
Gigabit Ethernet. Also untaring kernel archive lasts about 5 to 10 times
more then on local disks. I've tested some configs with write behind,
io-threads, io-cache - I can achieve only few percents better
performance.

I've also notice, that while writing small files (<=1kB) for every one
send packet there is one ACK recieving from server. I have tought that
with window-size option from write-behind translator I can get ACKing
for more packets then one. I thing there is a problem with low
performance, but that was my loosely ideas. ;)

Regards.

-- 
rash at konto pl