Having shot my mouth and hoping to extort some sushi the hardware vs
software raid testing should be happening this Sunday. It'll be like the
Super Bowl, but for sysadmin geeks.
The setup:
Dell 2850
2 x Xeon 3GHz w/2MB
6 GB RAM in 1GB sticks
5 x 73GB 10k U320
PERC4 inboard RAID card w/256 RAM
Gentoo 2005.1-r1 gentoo-sources 2.6.15-r1
By pulling the RAM on the PERC you turn it into a normal LSI Megaraid
card so the driver and hardware do not change between software and
hardware RAID. That should do a nice job of making sure that the way
we're doing RAID is the only defference per setup.
I have to look at the RAID card and see if it has a battery backup. If
not I'll leave write cache on for the disks and the card else I'll only
have the card write cache enabled... I may attempt both if there is
time. Software RAID will always have write cache on the disks. Towards
the end I'll try to do some tests on trying to break write caching on
both setups by pulling the power at inopportune times.
Cluster size is going to be 64k because that's the current sweet spot
and the default on the hardware RAID. If I get time I may also do 32k
for comparison. I'd want to try 8k as well in an attempt to throw more
data ops at the hardware raid in an attempt to swamp the processor, but
only an utter moron would actually use 8k clusters and I definitely
won't have time.
I may bump the RAM down to 1-2GB so I have less system cache and can
keep file sizes down. However I'd prefer to use the full 6 GB so that
the 256MB cache on the hardware RAID card is only 5% difference in total
RAM vs 25% in a 1 GB system. I may attempt to do both a 1GB and 6GB test
if only to see how the extra 256 MB affects things.
I'll only be testing RAID 5. That should be a good general case with
parity calculations, good read times, and decent write times so we can
push each setup fairly hard. I'll be using the same filesystem, ext3,
and using the same partitioning scheme in both cases. I will be using a
sixth drive as /boot for the software RAID since Linux can't boot RAID 5.
I'm still deciding the actual tests. bonnie, iozone, etc are going to
be in there, but I'm still looking for some other real world type tests.
I'm going to attempt to graph as much system data in each case so we can
plot CPU, ram usage, etc as well the actual data from the I/O tests.
I think that covers pretty most of what we should care about, but i'd
be open to suggestions of things I might have missed. Once testing is
done I'll pretty up the graphs, raw data, testing, setup, etc and post
it on a web page come Monday or Tuesday.
kashani
--
[email protected] mailing list
