latest PERC firmware == slow
John Vasileff
john.lists at gmail.com
Tue Jul 26 15:45:17 CDT 2005
For anyone using PERC raid that is concerned about IO performance, read on...
The latest firmware adds a "Patrol Read" feature that is supposed to
detect disk failures before they occur by reading from each disk that
is part of an array. By default, patrol read automatically runs every
4 hours (actually, 4 hours between each run.) On my test system,
patrol read takes about 20-25 minutes to complete.
According to the docs, patrol read is supposed to scale back during
periods of high IO, thereby minimizing impact. But, my tests show
severe impact to random IO. The first iozone test below is with
patrol read not running. The second is the same benchmark with patrol
read running. The entire first run was completed in about 10 minutes.
The second test didn't even finish the first random read test before
patrol read was complete - over 20 minutes.
### IOZONE when Patrol Read is NOT running
### 5 Disk RAID5
### (PERC 4e/Di, 2850, 2x 3.0GHz, kernel mem=512m)
# iozone -Ra -n 1G -g 1G -i 0 -i 2 -b output.wks
random random
KB reclen write rewrite read write
1048576 64 87756 86538 16940 30001
1048576 128 78120 87021 24289 42940
1048576 256 75545 89749 26922 48614
1048576 512 75519 91079 44636 66116
1048576 1024 78189 84996 70692 74730
1048576 2048 77202 84699 103527 71719
1048576 4096 61583 78150 130275 73023
1048576 8192 71793 78952 162446 70795
1048576 16384 71136 82325 185971 76474
### IOZONE when Patrol Read IS running
### The first "random read" test took over 20 minutes, Patrol Read was
### finished prior to the tests competion.
### (PERC 4e/Di, 2850, 2x 3.0GHz, kernel mem=512m)
### 5 Disk RAID5
# iozone -Ra -n 1G -g 1G -i 0 -i 2 -b output-pr.wks
random random
KB reclen write rewrite read write
1048576 64 71802 86163 715 25764
1048576 128 77812 94691 23645 43401
1048576 256
iozone: interrupted
John
More information about the Linux-PowerEdge
mailing list