Subject: other mem-related issue;was: overfilling mfs partitions
To: None <port-macppc@netbsd.org>
From: Rudi Ludwig <rudihl@gmx.de>
List: port-macppc
Date: 11/21/2004 13:59:13
Maybe unrelated, but then again I wouldn't post as a replay to this
thread.
The example of a user-program that cannot alocate the memory it wants
but that is obviously available (hardwarewise).

It is a small memory benchmark called HINT. Unfortunatly broken in
pkgsrc as of 2004Q3 but it is fairly simple to repeat and requires no
installation.

To repeat:
- Download the source at 
http://www1.physik.tu-muenchen.de/lehrstuehle/T35/Rechnerdoku/hint-2003-11-03.tar.gz
- untar
- cd hint-2003-11-03/unix ; make ; ./run.sh <hostname>
Runs a view minutes and will store the results in the folder
./data<hostname> (therefor run.sh needs a name to run)

The issue is:
- The benchmark increases problem-size with each iteration requireing
more memory. So it will show the L1-L2 .. RAM, SWAP boarders.
- It is designed to stop when its meassure drops to 10 percent of the
peak value, which should under normal conditions be when the machine
starts swapping (the value can be adjusted if needed).
Since integer and float exhaust numerical precision befor it gets that
far only the double results are of interest.

And here is the problem:
on macppc it doesn't get there, it stops with a memory related error
message in double.out
.............................................................ERROR

This run was memory limited at 1550648 subintervals -> 103466916 bytes

I am running
$ uname -a
NetBSD ibook 2.0_RC4 NetBSD 2.0_RC4 (IBOOK) #3: Mon Nov  1 19:46:38 CET
2004  rudi@ibook:/usr/src/sys/arch/macppc/compile/IBOOK macppc

but if I remember correctly, saw the same under macppc 1.6.2 , too.
Oh and physical RAM is 384MByte on the machine, so it doesn't even get
close to swapping.


Rudi