Dell/Perc5 raid/MPT SAS Integrated Raid Write Performance

mjacob at mjacob at
Fri May 18 23:31:04 UTC 2007

A lot of users have seen some very poor write performance on the MPT 
(LSI-Logic) driven SAS/SATA controllers, particularly those that 
function in Integrated Raid (mirroring) mode.

Some of the reported performance issues are pretty clearly single 
spindle small transfer IOPS issues. For example, directory intensive and 
small file operations like mail server applications can do very poorly 
on single spindle SATA drives that are connected via a SAS channel that 
doesn't enable write cacheing on the SATA drive (i.e., does not flow 
through WCE for SCSI emulation, as the LSI-Logic *apparently* does not). 
Benchmarks like Postmark show pretty amazing differences when run on a 
PATA or native SATA based drive (1000s of ops/second) and on single 
Fibre Channel or SCSI drives (100s of ops/second) and can be even worse 
for SATA drives on a SAS controller.

In these cases, there isn't much to be done- the h/w being picked 
doesn't match the application.

However, other users have reported things which are *clearly* bad 
performance issues. In these cases users have reported sequential write 
speed to be a small fraction of read speed. That is, a single threaded 
read of a 10GB file will get spindle rotational speed magnitude for the 
disk in question (~40MB/s) but will only write at around ~6MB/s. This is 
clearly broken and wrong.

Since I don't actually have a *lot* of MPT h/w and none that shows this 
write performance problem could folks do me a favor and at the next 
reboot get into the LSI-Logic BIOS utilities and find me all the 
firmware revision numbers? This might help me nail down some differences 
to go talk to LSI-Logic about.

The overriding LSI-Logic BIOS revision is of interest, but also any of 
the firmware revision numbers. For example, the loaned Sunfire 4100 I 
have has for the BIOS revision, but for the 
firmware- and this system, which has two integral SAS 2.5" drives, 
writes at 50MB/s with them set up as an integrated mirror.


More information about the freebsd-performance mailing list