Subject: Re: SCSI performance tuning on 5000/{125,133}, 5000/25
To: None <port-pmax@NetBSD.ORG>
From: Toru Nishimura <>
List: port-pmax
Date: 04/15/1997 14:35:37
About SCSI performance issue;  

> On Monday I committed a fix for a performance bug on the 3MIN
> (5000/125, 133) and MAXINE (5000/20,25,33) DECstations.
> The ioasic-attached 53c94 SCSI chip on these machines was being
> initialized for a 12.5 MHz clock speed (correct for a TurboChannel
> option card), when they should have been initialized for a 25MHz clock
> speed.
> ...
> Together, these changes should put the SCSI performance on a 3MIN or
> MAXINE back to what it was in pre-1.2 days.  I'd be interested to hear
> iozone or bonnie comparisons versus 1.2C, from those who reported a
> decrease in SCSI throughput since 1.2 or before.

I recompiled 1.2D NetBSD/pmax from 1997-4-14 tar_files, and tried it
with DS5000/25.  The result is fine.  With recent stock snapshot kernel
(seems compiled at 1997-3-15 PST), typical iozone performance was;
        337542 bytes/second for writing the file
        1154310 bytes/second for reading the file

And with recompiled 1.2D;
        1313840 bytes/second for writing the file
        1223685 bytes/second for reading the file

I had no time to measure old NetBSD/pmax (prior to asic_*.c split in
kernel tree) I/O performance this time.

Tohru Nishimura
Nara Institute of Science and Technology