benchmarking SCSI host adapters

[Piercarlo Grandi]

On 9 Nov 90 04:09:11 GMT, chip at chinacat.Unicom.COM (Chip Rosenthal) said:

chip> I have in my hot little hands two SCSI controllers - an Adaptec
chip> AHA-1542B and a BusTek BT-542B.  I've been trying to do some
chip> benchmark comparisons of the two using the "bonnie" program.
chip> [ ... ]

chip> I was told the BusTek gives about 3X the performance of
chip> the Adaptec.  I'm not seeing this - I see 1X performance.

	[ ... around 300KB/sec per block IO, read or write ... ]

chip> For reference, these were run on SCO ODT 1.0 (with no kernel
chip> tweaking).

Ahem. SCO Unix uses a fast file system of some sort, but from the
numbers you show I would guess that either the SCSI driver or the
filesystem are not as efficient as they could. I have seen better
numbers for the Adaptec than those you give.

chip> how useful is "bonnie" for disk system benchmarking?

Totally useless. "bonnie" tests for IO system performance -- the disk
subsystem is only the back end of an IO system. It is difficult to
abstract the performance of the disk subsystem from that of the whole IO
system, and probably pointless.

chip> Right now, my guess is that either I was given marketing performance
chip> numbers for the BusTek card rather than real performance numbers, or any
chip> additional performance of this card is just wasted potential on this 20MHz
chip> machine.

Unless your IO system sw is CPU bound (and this is quite difficult, even
if not totally impossible), and your numbers show that doing block IO
CPU intensity is 40-50%, CPU speed is not that important with an
intelligent controller. Setting it up properly and having an efficient
driver and filesystem implementation are far more important as limiting
factors.
--
Piercarlo Grandi                   | ARPA: pcg%uk.ac.aber.cs at nsfnet-relay.ac.uk
Dept of CS, UCW Aberystwyth        | UUCP: ...!mcsun!ukc!aber-cs!pcg
Penglais, Aberystwyth SY23 3BZ, UK | INET: pcg at cs.aber.ac.uk