Disk IO benchmarking

[Bill McGonigle]

On Jan 17, 2008, at 11:43, Ben Scott wrote:

>   I typically prefer hardware RAID for the same reasons.  Especially
> booting under fault conditions.  I can never really bring myself to
> trust that with software RAID, given all the weird problems I've seen
> with boot loaders over the years.

It's best to boot from RAID-1, so if you do have these problems you  
can boot off a drive as if it weren't RAID, and then rebuild the pack  
when your hardware is straightened out.  Software RAID is also good  
for when your RAID controller lets the magic smoke out on Saturday  
afternoon at 4, since there's little to no commonality among RAID  
headers from various manufacturers.  Where hardware RAID is important  
I spec an onsite spare for this situation (an onsite support contract  
with a good SLA would be another way to do this).  Otherwise we run  
software RAID-1 and change hardware easily.

It's also much faster on the gear I've been using.  I tried setting  
up a server recently with a 3Ware 9550SX-8 with hardware mirrors, for  
easy swap-n-replace for offsite backup.  This turned out to be far  
too slow for the system to be useful.  I re-did the drives as 8  
'single-unit RAID' volumes (JBOD doesn't seem to work for BIOS) and  
used Linux software RAID and got something like 6.5x performance  
boost, according to IOZone.  This was on a Xeon, at 2.66/1.33GHz, IIRC.

I'm also starting to compile IOZone files for flash sticks.  I wish  
there were a nice way to normalize these among hosts, because then an  
online repository would be possible, but I don't think that  
technology exists yet, so for now I'm just benching them all on the  
same hardware.

And I was incredibly dumb/tired and nuked those IOZone files from the  
3Ware as I rebuilt the mirrors.  I would have liked to have seen  
which kinds of operations were so slow (specific vs. uniform) but a  
deadline prevented re-running the tests.

Oh, yeah, IOZone can take a *long* time for a full test.  I've had to  
run it overnight on a slower (PNY) flash drive.  A comparable  
Kingston (price/capacity) was about 2-3 hours.  Windows has the  
[mumble mumble] flash test utility which claims to be better for  
flash testing, but I haven't seen one for unix yet, if the claim even  
has veracity.

-Bill

-----
Bill McGonigle, Owner           Work: 603.448.4440
BFC Computing, LLC              Home: 603.448.1668
bill at bfccomputing.com           Cell: 603.252.2606
http://www.bfccomputing.com/    Page: 603.442.1833
Blog: http://blog.bfccomputing.com/
VCard: http://bfccomputing.com/vcard/bill.vcf