Date:	Thu, 1 Jul 93 09:49:14 EDT
From:	DJ Delorie <dj AT ctron DOT com>
To:	LIP AT odie DOT ee DOT wits DOT ac DOT za
Cc:	djgpp AT sun DOT soe DOT clarkson DOT edu
Subject:	Re: efficiency: DPMI vs VCPI

> Has anybody benchmarked the speed of a DJGPP 1.10+ program running
> under VCPI and DPMI to gauge the relative efficiency of the two
> techniques?  The author of a PD app. built using the Intel 386/486
> code builder that I evaluated recently, claimed that his DPMI version
> was ~1.5x faster than the VCPI version. Any comments?

DPMI could be faster because it preloads the program instead of paging
it like VCPI does.  However, VCPI pages on demand and can handle much
larger programs.  It's a tradeoff.

> Secondly, the FAQ for DJGPP states (if I understand it correctly) that
> up to 128 Mb of memory and 128 Mb of virtual memory can be allocated
> giving a grand total of 256 Mb.  Is this a theoretical calculation or
> has this actually been tested and verified?  What was the PC setup
> that was used during the testing/verification (mem manager, config.sys
> and autoexec.bat)?

This is an upper limit; I know that go32 can't handle more than this
(but only by design - it's easily upped at the expense of some
conventional memory, about 4K per 128M).  The largest physical memory
I've seen used is a full 32Mb for number crunching; the largest swap
I've had is around 2-4Mb before I went out and got more RAM.  If your
XMS/EMS software can't handle more than 16M, then this will be your
upper limit.

DJ

- Raw text -