I/O performance problem on BL465c /SLES9 with SATA disks

[joseph pareti]

Has anybody had a similar problem? (see details below)
========================



- they compare dl385/U320 with bl465c and sas internal disks running sles9 sp3

- the write performance degrades of a factor 4x when doing a copy from one partition to another, while the read performance is OK

- the degradation is related to a driver under SuSE 9.3, and a rewritten driver on SuSE 10.0 fixes the problem

- they still need to run sles9

- on sles9 a special patch didn't fix (see below)

- I assume obvious things like battery write cache installed have been verified

- they use ext3.

=====================



The drivers attached with you mail didn't work on AMD64 architecture. No matter, I found them on the HP site.
As you can see on the tests below, results are not so good. With SAS disk and the new driver, performances depend on the cache efficiency. In development environment, the cache is not used more than 10% when datas are written on disks.

1. Tests results based on a copy of /usr
1.1. On servers with SCSI U320 disks

Hostname: ncegcolnx04
Server model: ProLiant DL145 G1
Disk model: SCSI 72Go U320
kernel version: 2.6.5-7.201-smp
Driver version: HP CISS Driver (v 2.6.6)
Load Average: 0.01
/usr size: 2506 MB
Total time: 104.64 seconds
write perf.: 23.949 MB/s

Hostname: ncegcolnx18
Server model: ProLiant DL385 G1
Disk model: SCSI 300Go U320
kernel version: 2.6.5-7.201-smp
Driver version: HP CISS Driver (v 2.6.6)
Load Average: 0.23
/usr size: 2455 MB
Total time: 57.05 seconds
write perf.: 43.576 MB/s

Hostname: ncegcolnx51
Server model: ProLiant BL25p G1
Disk model: SCSI 300Go U320
kernel version: 2.6.5-7.201-smp
Driver version: HP CISS Driver (v 2.6.6)
Load Average: 0.13
/usr size: 2600 MB
Total time: 54.18 seconds
write perf.: 47.988 MB/s


1.2. On servers with SAS disks

Hostname: ncegcolnx75
Server model: ProLiant BL465c G1
Disk model: SAS 146Go
kernel version: 2.6.5-7.244-smp
Driver version: HP CISS Driver (v 2.6.16-5)
Load Average: 3.13
/usr size: 2422 MB
Total time: 261.18 seconds
write perf.: 9.273 MB/s


1.2. With new driver given by HP

Hostname: ncegcolnx75
Server model: ProLiant BL465c G1
Disk model: SAS 146Go
kernel version: 2.6.5-7.244-smp
Driver version: HP CISS Driver (v 2.6.18-10)
Load Average: 0.20
/usr size: 2422 MB
Total time: from 340.64 to 23.51 seconds (cache efficiency)
write perf.: from 7.110 to 103.019 MB/s (cache efficiency)


2. Script used for these tests

#!/usr/bin/zsh
#2008-04-04
server_name=`hostname`
model=`hwinfo|grep 'Product: "'|cut -d: -f2`
kernel=`uname -r`
driver_version=`strings /lib/modules/$kernel/kernel/drivers/block/cciss.ko | grep 'HP CISS'`
load=`cat /proc/loadavg|cut -d\ -f1`
echo "
Hostname: $server_name
Server model: $model
kernel version: $kernel
Driver version: $driver_version
Load Average: $load"
size_m=$(( `du -sk /usr|awk '{print $1}'`/1024 ))
/usr/bin/time -o /tmp/time_result -f "%e" cp -rp /usr /data/usr_copy_test
time_s=`cat /tmp/time_result`
m_s=$(( $size_m/$time_s ))
echo " /usr size: $size_m MB
Total time: $time_s seconds
write perf.: $m_s MB/s"
rm /tmp/time_result
rm -r /data/usr_copy_test