Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 1 Nov 1996 22:24:37 +1100
From:      Bruce Evans <bde@zeta.org.au>
To:        dkelly@HiWAAY.net, jgrosch@superior.truenorth.org
Cc:        current@freebsd.org, jgrosch@sirius.com, jlemon@americantv.com, michaelv@MindBender.serv.net, roberto@keltia.freenix.fr
Subject:   Re: SCSI and IDE (was Re: 2.1.5r -> current upgrade)
Message-ID:  <199611011124.WAA07303@godzilla.zeta.org.au>

next in thread | raw e-mail | index | archive | help
>On the IDE drive:
>PeeCee: {872} time iozone 32
>[snip]
>IOZONE performance measurements:
>        1188424 bytes/second for writing the file
>        2111586 bytes/second for reading the file
>0.2u 10.8s 0:44.54 24.7% 69+750k 374+542io 0pf+0w
>
>Now on to the SCSI drive:
>PeeCee: {873} cd /usr1
>PeeCee: {874} time iozone 32
>[snip]
>IOZONE performance measurements:
>        3455323 bytes/second for writing the file
>        4149726 bytes/second for reading the file
>0.2u 10.9s 0:18.17 61.5% 26+281k 363+533io 0pf+0w
>     ^^^^ this is the interesting part.
>
>Notice how both took the same amount of system time?

This is normal.  IDE i/o is done in the interrupt handler
except for the first block, and interrupt overhead is not
counted in the system time.  SCSI i/o is done while the
CPU is not using the bus, and overheads for keeping the
CPU off the bus are not counted in the system time (except
indirectly).  In this benchmark, the parts of the overheads
that get counted are apparently almost identical.  I
estimate that the totoal SCSI overhead is 1-2 seconds and
3/4 of it gets counted (indirectly), while the total IDE
overhead is 10 seconds and 1/8 of it gets counted.

>I'm disappointed. A bit.

Use systat to see interrupt overheads, or just divide the
average transfer speed by the PIO speed and add a few percent.
E.g., PIO mode 0 (3MB/sec) delivering 2MB/sec has an overhead
of > 66%.  PIO mode 4 (16MB/sec) delivering 5MB/sec has an
overhead of > 30%.

OTOH, the large system overhead defeats half the advantage of
SCSI.  It seems a bit too large.  I get an overhead of about
2 seconds per 32M on a Triton 1 P133 system with both SCSI
and IDE drive delivering 5MB/sec.

Bruce



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199611011124.WAA07303>