From: Charles Sandmann <sandmann AT new-orleans DOT NeoSoft DOT com>
Subject: Re: Disk I/O rates with DJGPP
To: S DOT D DOT Smith AT ecs DOT soton DOT ac DOT uk (Sean Smith)
Date: Wed, 2 Nov 1994 08:48:10 -0600 (CST)
Cc: djgpp AT sun DOT soe DOT clarkson DOT edu, jnc AT ecs DOT soton DOT ac DOT uk

> A colleague of mine who is also using DJGPP has commented that the disk I/O
> (low level) for DJGPP objects is substantially slower than native 
> Windows or DOS I/O. The comment is that e.g. Zortech C/C++ can produce
> binaries that give block reads and writes of large blocks at up to
> 400kbs-1 while despite the chosen blocksize DJGPP low level reads and
> writes reach a ceiling at around 30kbs-1. Obviously there is some
> h/w dependency here, but I guess the ratios are comparable whatever the
> hardware.
> 
> I'm just off to look at the sources for read/write now, but does anyone
> have any comments on this?

I did some rather extensive benchmark testing in V2.x development on this
subject (both with and without disk caches).  I compared TCC, low-level
DOS, and DJGPP V1.x and V2.x.  A bug fix which improved V1.11 throughput
was implemented in one of the V1.11maint releases.

The short answer is our transfer buffer is too small.  This is not
changeable in V1.x but can be stubedited in V2.x.  I was able to read 
over 1Mb/sec into DJGPP with V2.0 code on large contig files on a 32Kb
transfer buffer (on a real disk, faster on RAM disk).  This was actually
about the same speed as the TCC or low level dos at that point.

If you really want speed, you can use the _go32_dpmi calls to allocate
your own 32Kb real buffer and do the reads manually.

BTW, the 30Kb/sec mentioned above must be on a text file on an uncached
drive, this is too low.