Some time ago I sent riel and quintela a program which, used with the right
parameter, turns into a kinda "worst case" memory trasher. Since it looks like
they didn't have enough time to look into it, I'm sending it to the whole list
now.

I also figured out the ideal number to stress test everything is 16 times your
memory size in bytes (of course you can use more or maybe less, which makes the
effect in the memory subsystem a bit different).

----- Forwarded message from cesarb -----

Date: Sun, 20 Aug 2000 01:25:34 -0300
To: quintela@fi.udc.es
Cc: riel@conectiva.com.br
Subject: [cesarb: sieve.c]
User-Agent: Mutt/1.2.5i

riel thinks this should be included in your memtest suite. So I put it in
public domain (as something that simple deserves).

It was a sieve program I made last year. Was also my first pthreads program
(since I *had* to move the printf out of the main loop, it was killing the
performance).

riel says it is a pretty realistic sweep pattern.

The only change you might want to make to it (to make it work even better as a
VM test) would be to comment out the printf ("%u ", I2N (i)); from the output
loop, so the last sweep still gets done but is not dependent anymore in the tty
speed (you should also comment out the printf ("%u ", 2); line). Leave the
final count output there to check for bad runtime errors.

If you want to use it for even larger numbers (but getting a bit more CPU
bound) you should comment out the output lines I mentioned above, change bitpos
to a unsigned long long, and change BITPOS_FORMAT to match. I didn't test that,
but the program was ready for that kind of change from day one.

You could also include this email in your test suite's documentation. It's
pretty much the only documentation sieve.c would get.

----- Forwarded message from cesarb -----

Date: Sun, 20 Aug 2000 00:57:32 -0300
To: riel@conectiva.com.br
Subject: sieve.c
User-Agent: Mutt/1.2.5i

Here is it. I compile it with gcc -W -Wall -O3 -march=k6 -lpthread -save-temps

You need 1 billion (bah, screw the english. Um bilhao) maximum to get 60Mb
core. 2 billion get you 118Mb, which causes severe trashing and awful latency
on a 128Mb machine with X+gnome+mozilla+xmms+xchat+lotsa xterms+the kitchen
sink running.

----- End forwarded message -----

----- End forwarded message -----

-- 
Cesar Eduardo Barros
cesarb@nitnet.com.br
cesarb@dcc.ufrj.br