CERC RAID5 performance lower than everyone else

[Adam Nielsen]

> Just out of interest I thought I'd try a test to see what difference it made
> on our CERC.  I'm not drawing any conclusions from the results ;)

I think it depends on what task you run.  Using mkisofs to create a 
531MB .iso I get this:

--setra  Read/write   cat > /dev/null
64       9.8 MB/sec   32.7 MB/sec
128     10.9 MB/sec   46.4 MB/sec  (Slackware default)
256     11.1 MB/sec   58.0 MB/sec
512     11.7 MB/sec   69.5 MB/sec
1024    13.0 MB/sec   77.9 MB/sec
2048    12.6 MB/sec   102.0 MB/sec
4096    13.9 MB/sec   111.9 MB/sec
8192    14.3 MB/sec   112.1 MB/sec
16384   14.8 MB/sec   110.7 MB/sec
32768   15.6 MB/sec   117.0 MB/sec
65536   18.2 MB/sec   120.8 MB/sec

The figures are a bit rough as there seems to be quite a bit of 
variation, so this is the average of three runs.  The read/write figure 
is the time taken to run mkisofs divided by the image size, so for 
--setra 64 it means it averaged at 9.8MB/sec read white simultaneously 
averaging 9.8MB/sec write - of course the actual speeds were quite 
different, but this is just a comparison after all.

Likewise the "cat" column is how fast it read a different 500MB file, 
primarily to clear the disk cache before the next run of mkisofs.

Like I say these figures are quite rough, but it does seem to indicate 
there is at least some improvement by increasing the readahead.  At the 
higher end I was actually seeing the transfer peak at just over 
170MB/sec, so I'm surprised that something as simple as readahead could 
make such a difference.

Cheers,
Adam.