PE2950, LSI SAS -> SATA very slow

Adam Nielsen adam.nielsen at uq.edu.au
Tue May 4 19:57:39 CDT 2010


>>> 10MB/sec does seem extremely poor. we've had read slowness when Linux block
>>> device read-ahead was too small. Once properly configured, performance on
>>> RAID-5 went from 150MB/sec to 500MB/sec, so the difference is dramatic. we
>>> usually tune read-ahead (in Linux) buffers starting from 8MB to 32MB and
>>> benchmark to see what works best. but your problem may be elsewhere... we've
>>> never seen 10MB/sec...
>> Have you got any pointers to where this can be adjusted?  It seems 
>> (according to Google) there are countless methods and patches, most of 
>> which are for rather old kernel versions.
> 
> man blockdev
> blockdev --setra blocks /dev/sda
> 
> It is part of util-linux-ng .

Thanks for the hint!  This didn't make much difference for us, but 
that's probably because the array is RAID-1, so the array readahead 
already matched the individual disk readahead.  Strangely enough 
increasing the readahead actually reduced the raw speed according to 
hdparm -t, which I wouldn't have thought would make a difference.

What did make a huge difference is setting the NCQ depth to 1 (instead 
of the default 64) as Tim suggested.  This brought raw read speed up 
from ~50MB/sec to almost 90MB/sec!

Haven't yet had the machine under a large load so will be interesting to 
see how it copes, but now I really want the updated firmware from Dell!

Cheers,
Adam.



More information about the Linux-PowerEdge mailing list