Subject: Re: ciss driver HP Smart Array 5i
To: SODA Noriyuki <firstname.lastname@example.org>
From: Joseph A. Dacuma <email@example.com>
Date: 05/21/2006 14:17:01
Hello Soda San!
>> Hi All!
>> Anyone on the list using the ciss driver? Its being tested on an HP
Proliant DL380 G3 server with a Smart Array 5i controller. Preliminary
findings reveal that I/O operation is very slow. This applies to both
>> 1 configuration using 2 disks and RAID 5 using 6 disks.
>> Due to its crawling read/write performance, I tried to install OpenBSD 3.9
>> on the same machine using same configurations above. To my surprise, ciss
>> driver of OpenBSD 3.9 performs better on both RAID 1 and 5 --albeit, twice
>> the speed of I/O operations compared to NetBSD-3 branch.
> Mm, that surprises me too.
Same here, I've always thought (my experiences as well) that NetBSD was
faster than OpenBSD. But in this case its not. :)
>> Using generic kernels for both variants, simulation using time was
utilized to compare the I/O speed difference by upacking an 8.8MB
>> file from CD. It took 28:41.22 for NetBSD. On OpenBSD 3.9, the
operation was completed less than half the time clocking at 14:25.03.
> Both sounds somewhat slower on a machine today.
> On my machine (NetBSD-3.0, Athlon 64 X2 2.2GHz, 2GB memory, 160G IDE
disks, ffs without softdep), unpacking ghostscript-8.51.tar.bz2 source
tarball only took 5.45 seconds (on 7200RPM disk) or 6.07 seconds (on
5400RPM disk), although the source tar ball is from a harddisk instead
of a CD.
I wanted to check each and every aspect of the machine since its the first
time NetBSD ran ciss on HP/Compaq servers. I'm very optimistic that this
will increase the number of NetBSD user base as there are many of these
machines deployed in the world.
>> Also, while unpacking the tar.gz file, I noticed that stdout pauses
invariably after 2-5 screenful of outputs. On OpenBSD's stdout, files,
while being unpacked flowed smoothly as entire operation ended.
> Are you using softdep on both machines?
> Does changing vm.bufcache parameter by sysctl make any difference? --
I only do those adjustments in tiny increments after successfuly deploying
it in production. Also, I'm no longer using softdep on my machines as I
put more importance on stability and data integrity than speed. Not that
softdep is dangerous to use but I don't want leave anything to chance when
it comes to machines deployed for production.
On both machines, It was just a default install, no changes whatsoever.
IMHO, there should be no tweaks to have the machine run smoothly on