PERC 3/DC - Dual Channel performance

Sigbjorn Strommen sigbjorn.strommen at
Mon Jun 10 16:17:01 CDT 2002

Seth Mos wrote:
> At 19:32 10-6-2002 +0200, Sigbjorn Strommen wrote:
> >Well, if you're going from 3 times 2 disks in raid 1 to 2 times
> >3 disks in raid 1, you will get less disk space available.
> 3 disks in raid 1 ???

Don't know why (but that's how I read the original email)...

> Or 2 disks in raid1 with one hot spare.
> >In addition, the performance increase in using two channels vs.
> >one channel is not that great with the PERC's, as they are just
> >too slow to start with.
> >I've done tests where I shuffled data between disks on different
> >channels, and it turned out to be just a few percents faster then
> >doing the same tests between partitions on one disk...
> >Even raid 0 across the two channels did not help.
> I did some bonnie++ benchmarks with software raid10 vs hardware raid10. The
> biggest jump is in the rewrite speed. And larger chunks then 128K also help
> a lot.

Possibly.  I tried chunksizes 32, 64 and 128, with no differences
(on real data at least).

BTW, I notice you get a lot higher block reads then I did (well, all
your numbers looks better actually), maybe you need to increase your
data size a bit (from my tests with bonnie++ it looked like 2 times
the RAM is not enough)

What kind of card is the "LSI logic Megaraid 473"?  I thought the
PERC 3DC's where the same as the Elite 1600?

> >Conclusion is, the PERC's are slow whatever config you are setting
> >up, and there's nothing you can do about it except for hoping that
> >future drivers and firmware will help a little (but even so, there's
> >nothing you can do about the slow processor on the controller).
> You will especially notice this with raid 5 if you have a lot of disks.

Funny thing with my setup (PE2500 with 1*PERC 3Di and 2*PERC 3DC) was
that I got *better* results with 5 disks in raid 5 on one channel then
4 disks in raid 0 across two channels, or 3 disks in raid 0 on one
channel, etc...
...kind of strange...

> The controller can handle 4 disks just fine but software raid is already
> showing it's faster in most respects.

Well, in my tests it's not only faster.  It's MUCH faster, and it
handles higher loads better!  In one test I executed 8 bonnie tests
simultaneously on 4 scsi channels (2*39160) and software raid 5.
If my memory serves me *all* the different categories in bonnie++
turned out better results for each of the bonnie instances then a
single bonnie test against any raid config on the megaraid...

> What I don't understand is that the rewrite speed of raid5 is higher then
> with hardware raid 10. On both occasions I was using cached IO on the
> controller and 128KB chunk size.

Well, that would fit with my results as well.


More information about the Linux-PowerEdge mailing list