And the long hang do happen... once you start getting a bit of memory
pressure
(say you go from 7000 to 7200 VMs and you only have memory for 7150) then you
are hitting the long delays *for every page* the VM inspects, and it will
I don't understand "*for every page*": why for *every* page?
I won't dispute "for many pages, many more than is bearable".
for every page the VM inspects; the VM tries to free memory and goes on trying
to free stuff, but (I'm guessing here) skipping active pages, but to know and clear
active, you need to walk the whole chain. for each page.
Now, you can make it 2x faster (reboot in 12 hours? ;-) ) but there's really
a much
higher order reduction of the "long chain" problem needed...
I'm with Andrea that prevention of super long chains is the way to go, we can
argue about 250
or 500 or 1000. Numbers will speak there... but from a KSM user perspective,
at some point
you reduced the cost of a page by 250x or 500x or 1000x... it's hitting
diminishing returns.
I'm not for a moment denying that there's a problem to be solved,
just questioning what's the right solution.
The reclaim case you illustrate does not persuade me, I already suggested
an easier way to handle that (don't waste time on pages of high mapcount).
Or are you saying that in your usage, the majority of pages start out with
high mapcount? That would defeat my suggestion, I think, But it's the
compaction case I want to think more about, that may persuade me also.
well in most servers that host VMs, of the, say 128Gb to 240Gb, all but a few hundred MB
is allocated to VMs, and VM memory is generally shared. So yes a big chunk
of memory will have a high map count of some sorts.
--
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>