Nice test, it's simpl and fairly easy to analyze. It provides a nice mix of
nonlocalized and localized memory activity, in that order.
The first pass of the quicksort will make one pass over all of memory with a
working set within the array at any given time of just two pages,
corresponding to the pattern of exchanges. Somewhat more than the available
physical memory will be dirtied, causing some swapouts. A prescient mm would
page back in each of those swapped out pages exactly once later in the sort,
by choosing all the eviction victims from the set of pages that will not be
used in the next recursive semi-sort. I can't imagine any general-purpose
page replacement policy that would be this smart, so we should assume that a
good page replacement algorithm will evict about 50% more dirty pages than
the prescient algorithm.
There is some danger that any given page replacement algorithm might
accidently behave the same as the prescient algorithm. I can't think of a
good way to test for this other than by knowing how much physical memory is
actually available for the array and becoming suspicious if the mm does
better than expected according to the above analysis. Odds are, the mm won't
get lucky, but it could.
Aside from obvious bugs, one mm stategy can beat another by making more
memory available for caching the array data and by closely approaching 150%
of the number of swapouts that the prescient algorithm would make.
Andrea's mm apparently did better than Rik's, but you were too polite to
point that out ;-) Some additional statistics would help tell us why:
swapin/swapout counts, and run time with no swapping, which you would have to
obtain using a smaller data size and an artificial memory limit. It would
also be nice to know the maximum array size that can be sorted with no or
minimal swapping, which tells us how much cache the mm is able to make
available for program data. Ideally, that would amount to nearly all of
physical memory since very little else is going on.
-- Daniel - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to firstname.lastname@example.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/