Newsgroups:	comp.os.msdos.djgpp
Subject:	Re: large variation in run-time
Message-ID:	<PSxEvgVLkg3c@ns2>
From:	"Johan Van Ooijen" <j DOT w DOT vanooijen AT cpro DOT dlo DOT nl>
Date:	Thu, 7 May 1998 09:53:13 +0100
References:	<354DAFD3 DOT 8D277F06 AT CPRO DOT DLO DOT NL> <6iktlt$4j1 AT argo DOT telecom DOT cz>
NNTP-Posting-Host:	10.194.2.147
Lines:	46
To:	djgpp AT delorie DOT com
DJ-Gateway:	from newsgroup comp.os.msdos.djgpp

Eli Zaretski wrote:

>On Mon, 4 May 1998, Johan van Ooijen wrote:
>
>> So the question is, whether there is an explanation for these large
>> variations in run-time when using go32 ?
>
>Does the program use disk I/O significant proportion of its run time?

No, once every millions of iterations a 100 lines are written to a file and
one line to the screen.

>Does it swap to disk?  If so, your disk cache can skew the run time a
>great deal.

I don't see swapping activity. Memory allocated before the iterations is
less than 100 Kb, in the iterations about 1 Kb is malloc()'ed and free()'ed
each time. Total memory allocated doesn't grow. There's sufficient megabytes
of free memory.

>Another possibility might be the size of your innermost loop: if it
>sometimes overflows the CPU caches, you will see a factor of 2-4 in
>performance without any apparent explanation.

How does this happen?

>> and go32 says:
> >   go32 version 1.12.maint3
>
>Huh?  Is this program compiled with DJGPP v1.12?  If it's v2.01, run
>go32-v2, not go32 (it's a completely different program, and reports
>completely different results).

Yes, I still use go32 version 1.12 and gcc version 2.6.0. It works great for
me (except for this strange phenomenon...). I'll go over to go32-v2 some
time...
By the way, I also ran these simulations on an Alpha under OpenVMS using
dec-cc and there everything runs as expected.

Johan van Ooijen
j DOT w DOT vanooijen AT cpro DOT dlo DOT nl

- Raw text -