Klaus Schroer <[EMAIL PROTECTED]> wrote:
> This suggests that the DPT controller itself is not very efficient and
> that even the Megaraid cannot beat a single controller software raid setup
> (at least if you only use 3 disks). It might be a different stroy if you
Yes. However you will note that a true hardware RAID controller will
normally lower your CPU usage compared to a software RAID. However,
main CPU is nowadays far cheaper than specialized CPUs like you can
find on hardware RAID PCI adapters ...
> use more than four disks and thus saturate the scsi bus. The read values
> are what you would expect (3 disks = 2*single) but the write overhead
> seams to be significant. Is that to be expected?
Yes. For RAID5, you need to remember two things:
1. if you have three disks, read performance will at best be
2 x one disk, for one streaming sequential I/O
2. write performance may be lower, since to write N kilobytes
of data you need to write 2*N/2 to each disk, compute the
parity (CPU intensive), and then write N/2 to the current
parity disk (RAID5 has rotating parity). So, N becomes 3N/2
kilobytes.
> With a four disk setup you get a significant performance increase for
> writing but not very much for reading (again software raid with one
> AHA2940U2W LVD):
Are you sure you run U2LVD ? Look at the dmesg output for the aic7xxx
driver. Is there an archive for linux-raid somewhere ? I posted a few
benchmarks made on upto 8 Seagate Cheetah with RAID5 and RAID0, changing
things such as splitting in many host adapters, etc. It looks like
you can't get very good performance on U2LVD with more than three
Cheetah per channel.
> 1*1024 5455 90.9 11316 17.6 3626 11.7 5196 74.6 10308 13.9 128.9 4.2
What disk is that ? I was thinking the 18 GB IBMs were 10'000 RPM, around
18-24 MByte/s. This looks more like a 7200 RPM drive of older generation
(IBM DDRS can do 12-13 MByte/s).
Or it could be fs overhead: are you using mke2fs -b 4096 ?
Or it could be that your partition is at the end of the disk ?
PS: how much memory do you have ? With 256 MB memory I like to benchmark
with 2 GB files.
