i haven't tried all of arca8 yet, but i did sniff around in the bdflush
logic to see what you tried out. i took a stock 2.2.5 kernel and took out
the "run_task_queue()" in wakeup_bdflush() just to see what that it would
do. that made things a little worse, so i'm guessing you also had some
compensatory changes in other areas.
i'm interested to see how your flushtime changes will effect performance.
> in arca8 I restored the o+offset in the hash function, feel free to remove
> it before running the bench if you think it will harm (the +offset is not
> needed until you'll be very low on memory and it looks like to me that
> your machine never gets low on memory ;).
i'm discovering that a 13 bit hash mitigates the spikey size distribution
in the page hash *better* than the +offset change. although i've been
able to push the system into swap, i still haven't seen any degenerate
hash behavior that's as bad as the buffer cache's hash function.
i'll have more as i test this further.
> Ah and also the first pagemap-cachealigned patch was wrong (blame me). The
> second patch (after Eric pointed out to think about it twice) does the
> right thing (so you may consider to run the bench on it again... excuse
> me). So probably this is the reason that the pagemap-aligned+irq-aligned
> bench was weird...
i suspected that the strange performance behavior wasn't an implementation
problem, so i tried a plain 2.2.5 kernel with just the page struct
alignment mod in your latest patch. the numbers were about the same as
before. it would probably be a good idea if someone could actually take a
close look at what the alignment mod is doing to change hardware behavior,
'cause it's obviously not what we expect.
- Chuck Lever
-- corporate: <chuckl@netscape.com> personal: <chucklever@netscape.net> or <cel@monkey.org>The Linux Scalability project: http://www.citi.umich.edu/projects/citi-netscape/
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/