From: ws@xivic.bo.open.de (Wolfgang Schelongowski) Subject: Re: SCSI Performance (Yet Again) Date: Wed, 25 Aug 1993 12:04:32 MEST
jhenders@jonh.wimsey.bc.ca (John Henders) writes:
> ws@xivic.bo.open.de (Wolfgang Schelongowski) writes:
>
> >How about
> > date;dd if=/dev/sda of=/dev/null bs=131072 count=512;date
>
> A more accurate measure of the scsi code, IMHO, would be to use code
> which would use scsi read and write instructions to do real low level
> i/o, removing all other variables.
That would measure the "bare metal", and most people would find it
difficult to do. What I proposed was measuring the "bare metal" plus
the lowest level SCSI-driver where everybody can easily get at.
Comparing that with the "bare metal" would give some clue where
the below mentioned speed is lost.
> >That took 69 seconds => 950 KByte/sec for raw I/O. Under BSDI,
> >the numbers are 36 seconds on /dev/rsd0a => 1820 KByte/sec.
>
> >As I'm still running 0.99 pl6 the numbers for Linux may have
> >changed (hopefully to the better). Can somebody test it and
> >post the results so we have _numbers_ to compare and not
> >assumptions ?
>
> >My System: noname 486/33, AHA1542B, Fuji M2624FA. It makes no
> >difference for Linux whether I use async or sync SCSI.
>
> I ran the above test and got 103 seconds, with a 386/40 and a Maxtor
> 120 HD. A test for scsi i/o should not be affaected by processor speed,
> should it?
No. But I forgot to mention that I have 256KB of CPU/Main-Memory
cache. And your drive (its probably called a Maxtor 7120S here)
has 15ms access time and only 1536 KByte/sec transfer speed.
According to the specs of my drive it can easily saturate the
ISA-Bus. And what is your patch-level ?