[geeks] Rack-mount SCSI...

[Greg A. Woods]

[ On Monday, August 27, 2001 at 21:33:05 (+0000), Kris Kirby wrote: ]
> Subject: Re: [geeks] Rack-mount SCSI...
>
> Better post a Bonnie run or I'll be forced to scream "IDE!" at the top of
> my lungs. Please don't make me do this. :)

You can't compare RAIDframe RAID-5 (i.e. host-based software RAID) with
IDE of any sort.  I'm doing parity calculations and RAID management in
the kernel with no hardware assist.

You also can't compare those model Quantum drives with anything current,
IDE or not.  They're not particularly fast as such things go.

I'm not even sure you can compare an Adaptec wide differential
controller (I forget the stupid model number, but it has a aic7880 chip)
with anything modern either (IDE or not).  The ahc driver in NetBSD's
also a full year behind the times w.r.t. the FreeBSD sources it's
derived from (and IIRC there's a linux driver derived from the same
FreeBSD driver too).

Finally note that I didn't tune this array and its filesystem for large
files (ala Bonnie), but rather more with zillions of smaller files in
mind.

With those qualifications in mind, and considering this particular test
is run with the machine in multi-user mode:

             -------Sequential Output-------- ---Sequential Input-- --Random--
              -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine    MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU  /sec %CPU
/home     100  4291 45.4  5497 20.0  4757 23.0  8052 74.3 64151 85.3 1951.3 74.5

Hmmm... probably should have done that with a bit bigger test file as
the block input speed is a factor of 5 too large (raw read speed from
the /dev/raid0a device is 12MB/s).  The other numbers are pretty much
what I got from a 1GB test file in single-user mode the other day
though.

Raw write speed through a freshly created filesystem with dd from
/dev/zero is a bit higher with 64kb blocks than the block output above,
probably mostly becasue the I/O's are one-to-one, but I think it may
also be because RAIDframe can then do the parity calculations more
efficiently (it may be optimised to calculate parity for long strings of
zeros for all I know).

-- 
							Greg A. Woods

+1 416 218-0098      VE3TCP      <gwoods at acm.org>     <woods at robohack.ca>
Planix, Inc. <woods at planix.com>;   Secrets of the Weird <woods at weird.com>