On (07/14/15 09:55), Minchan Kim wrote:
> > It depends on 'big overhead' definition, of course. We don't care
> > that much when compaction is issued by the shrinker, because things
> > are getting bad and we can sacrifice performance. But user triggered
> > compaction on a I/O pressured device can needlessly slow things down,
> > especially now, when we drain ALMOST_FULL classes.
>
> You mean performance overhead by additional alloc_pages?
not only performance, but yes, performance mostly.
> If so, you mean ALMOST_EMPTY|ALMOST_FULL, not only ALMOST_FULL?
of course, I meant your recent patch here. should have been 'we _ALSO_
drain ALMOST_FULL classes'
>
> So, it's performance enhance patch?
> Please give the some number to justify patchset.
alrighty... again...
> >
> > /sys/block/zram<id>/compact is a black box. We provide it, we don't
> > throttle it in the kernel, and user space is absolutely clueless when
> > it invokes compaction. From some remote (or alternative) point of
>
> But we have zs_can_compact so it can effectively skip the class if it
> is not proper class.
user triggered compaction can compact too much.
in its current state triggering a compaction from user space is like
playing a lottery or a russian roulette.
a simple script
for i in {1..1000}; do
echo -n 'compact... ';
cat /sys/block/zram0/compact;
echo 1 > /sys/block/zram0/compact;
sleep 1;
done
(and this is not so crazy. love it or not, but this is the only way
how user space can use compaction at the moment).
the output
...
compact... 0
compact... 0
compact... 0
compact... 0
compact... 0
compact... 0
compact... 409
compact... 3550
compact... 0
compact... 0
compact... 0
compact... 2129
compact... 765
compact... 0
compact... 0
compact... 0
compact... 784
compact... 0
compact... 0
compact... 0
compact... 0
...
(f.e., we compacted 3550 pages on device being under I/O pressure.
that's quite a lot, don't you think so?).
first -- no enforced compaction
second -- with enforced compaction
./iozone -t 8 -R -r 4K -s 200M -I +Z
w/o w/compaction
" Initial write " 549240.49 538710.62
" Rewrite " 2447973.19 2442312.38
" Read " 5533620.69 5611562.00
" Re-read " 5689199.81 4916373.62
" Reverse Read " 4094576.16 5280551.56
" Stride read " 5382067.75 5395350.00
" Random read " 5384945.56 5298079.62
" Mixed workload " 3986770.06 3918897.78
" Random write " 2290869.12 2201346.50
" Pwrite " 502619.36 493527.64
" Pread " 2675312.28 2732118.19
" Fwrite " 4198686.06 3373855.09
" Fread " 18054401.25 17895797.00
> > view compaction can be seen as "zsmalloc's cache flush" (unused objects
> > make write path quicker - no zspage allocation needed) and it won't
> > hurt to give user space some numbers so it can decide if the whole
> > thing is worth it (that decision is, once again, I/O pattern and
> > setup specific -- some users may be interested in compaction only
> > if it will reduce zsmalloc's memory consumption by, say, 15%).
>
> Again, your claim is performace so I need number.
> If it's really horrible, I guess below interface makes user handy
> without peeking nr_can_compact ad doing compact.
>
> /* Tell zram to compact if fragment ration is higher 15% */
> echo 15% > /sys/block/zram0/compact
> or
> echo 15% > /sys/block/zram/compact_condition
no, this is the least of the things we need to do -- enforced and
pre-defined policy engine in zram/zsmalloc 'that will work for all'.
user space has almost all the numbers to do it, the only missing bit
of the puzzle is `nr_can_compact' number. it's up to user space then
to decide how it wishes things to be done. for example:
"don't compact if compaction will flush 35% of zsmalloc pages on a
I/O pressured device" or something else.
-ss
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>