OSEC

Neohapsis is currently accepting applications for employment. For more information, please visit our website www.neohapsis.com or email hr@neohapsis.com
 
Subject: Re: SCSI, U2W in particular?, and poor performance...
From: Liviu Daia (Liviu.Daiaimar.ro)
Date: Sun Mar 05 2000 - 15:41:33 CST

On 5 March 2000, Dylan Griffiths <Dylan_Gbigfoot.com> wrote:
> It could be a problem with the Linux SCSI subsystem. While stable, is
> it somewhat evil code as of the 2.2.x kernel revisions. Much works
> has been done in the 2.3.x development series to clean up the SCSI
> subsystem (with good results).

    Since you mention it: this is a common confusion. The driver that
used to be a big mess and has been revised in 2.3.x kernels is the SCSI
_generic_ driver (/dev/sg*, as opposed to /dev/sd*), used for "SCSI
emulation" with IDE CD writers, DATs, Zip drives, parallel port discs,
and a few other oddball devices, but _never_ with "normal" internal
discs. The driver for internal discs hasn't change much in 2.3.x,
except for adding support for new controllers.

> Also, the SCSI HDs likely support TCQ (tagged command queing), but
> this is not enabled by default in the Linux kernel for compatibility.
> I suggest he tunes the Linux kernel SCSI AIC7xxx driver parameters
> before he does anything else, as the "most compatible" SCSI mode is
> slower than the "best speed" mode of IDE (which is also default).

    Now, this may be indeed the real crux of the matter. The design of
the PC DMA is simply abysmal; without TCQ the latency of a SCSI disc
on a PC can be as much as 40 times larger than the one you'd normally
get from the same disc on a less brain-dead machine, say a Sparc.
(Typically the throughput would be smaller too --- this time because
of the abysmal PC bus --- but that can't be "fixed" in software.) But
for some reasons whoever was supposed to document this for Linux either
didn't understand what this TCQ was about, or was a big fan of black
magic.

> Otherwise, this is not a postfix issue, nor does it nessitate further
> discussion/flaming on this list.

    Regards,

    Liviu Daia 

-- 
Dr. Liviu Daia               e-mail:   Liviu.Daiaimar.ro
Institute of Mathematics     web page: http://www.imar.ro/~daia
of the Romanian Academy      PGP key:  http://www.imar.ro/~daia/daia.asc