Path: utzoo!utgpu!news-server.csri.toronto.edu!rutgers!cs.utexas.edu!sdd.hp.com!ucsd!ucbvax!UCSD.EDU!sdempsey
From: sdempsey@UCSD.EDU (Steve Dempsey)
Newsgroups: comp.sys.sgi
Subject: strange floating point exception interrupt behavior
Message-ID: <9007260022.AA03729@chem.chem.ucsd.edu>
Date: 26 Jul 90 00:22:26 GMT
Sender: daemon@ucbvax.BERKELEY.EDU
Organization: The Internet
Lines: 40

The following discussion pertains to a 4D/25TG running 3.2.1 and a
4D/340VGX running 3.3.

Recently I have been doing a performance analysis of a number cruncher
program that runs much more slowly on IRISes than one would expect.
I fired up gr_osview and ran the program, expecting to see lots of
system calls or swapping, and an indication of where the cpu time was being
wasted.  What I saw was something quite strange!  The cpu was spending
99% of its time in user mode, just like any decent number cruncher should.
The shock came from the interrupt rate, which went from a background level
of 200-400 per second up to ~20K per second (35K on the 340VGX!)

Ultimately, I discovered that the extra interrupts were occuring whenever
floating point operations resulted in underflow.  This behavior can be
demonstrated by compiling and running this code:

    #include <math.h>

    main()
	{
	double x, y, z;
	int	i;

	y = MINDOUBLE;
	z = 0.5;
	i = 10000000;
	while(i--) x = y * z;
	}

Both C and Fortran versions of this code produce the same results.
I tried similar tests, forcing overflows and divide-by-zero, but no extra
interrupts were found for these floating exceptions.

Can anybody explain what's so special about underflows, and why do I get
interrupts even though the floating point exception interrupts are not enabled?
--------------------------------------------------------------------------------
Steve Dempsey						       (619) 534-0208
Dept. of Chemistry Computer Facility, 0314	   INTERNET:   sdempsey@ucsd.edu
University of California, San Diego		     BITNET:   sdempsey@ucsd
La Jolla, CA 92093-0314				       UUCP:   ucsd!sdempsey