VM tuning through fault trace gathering [with actual code]

From: John Fremlin (vii@users.sourceforge.net)
Date: Mon Jun 25 2001 - 10:26:39 EST


Last year I had the idea of tracing the memory accesses of the system
to improve the VM - the traces could be used to test algorithms in
userspace. The difficulty is of course making all memory accesses
fault without destroying system performance.

The following patch (i386 only) will dump all page faults to
/dev/biglog (you need devfs for this node to appear). If you echo 1 >
/proc/sys/vm/trace then *almost all* userspace memory accesses will
take a soft fault. Note that this is a bit suicidal at the moment
because of the staggeringly inefficient way its implemented, on my box
(K6-2 300MHz) only processes which do very little (e.g. /usr/bin/yes)
running at highest priority are able to print anything to the console.

I think the best way would be to have only one valid l2 pte per
process. I'll have a go at doing that in a day or two unless someone
has a better idea?




-- 
PS. I'm desparately seeking last minute summer job in Europe due to an
unfortunate series of events. Please see http://ape.n3.net/cv.html

- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/



This archive was generated by hypermail 2b29 : Sat Jun 30 2001 - 21:00:12 EST