On Sat, 18 Sep 2010, Robert Watson wrote:
On Fri, 17 Sep 2010, Andre Oppermann wrote:
Although keeping free items around improves performance, it does consume
memory too. And the fact that that memory is not freed on lowmem
condition makes the situation worse.
Interesting. We may run into related issues with excessive mbuf (cluster)
caching in the per-cpu buckets as well.
Having a general solutions for that is appreciated. Maybe the size of the
free per-cpu buckets should be specified when setting up the UMA zone. Of
certain frequently re-used elements we may want to cache more, other less.
I've been keeping a vague eye out for this over the last few years, and
haven't spotted many problems in production machines I've inspected. You can
use the umastat tool in the tools tree to look at the distribution of memory
over buckets (etc) in UMA manually. It would be nice if it had some
automated statistics on fragmentation however. Short-lived fragmentation is
likely, and isn't an issue, so what you want is a tool that monitors over
time and reports on longer-lived fragmentation.
Not specifically in reaction to Robert's comment but I would like to add
my thoughts to this notion of resource balancing in buckets. I really
prefer not to do any specific per-zone tuning except in extreme cases.
This is because quite often the decisions we make don't apply to some
class of machines or workloads. I would instead prefer to keep the
algorithm adaptable.
I like the idea of weighting the bucket decisions by the size of the item.
Obviously this has some flaws with compound objects but in the general
case it is good. We should consider increasing the cost of bucket
expansion based on the size of the item. Right now buckets are expanded
fairly readily.
We could also consider decreasing the default bucket size for a zone based
on vm pressure and use. Right now there is no downward pressure on bucket
size, only upward based on trips to the slab layer.
Additionally we could make a last ditch flush mechanism that runs on each
cpu in turn and flushes some or all of the buckets in per-cpu caches.
Presently that is not done due to synchronization issues. It can't be
done from a central place. It could be done with a callout mechanism or a
for loop that binds to each core in succession.
I believe the combination of these approaches would significantly solve
the problem and should be relatively little new code. It should also
preserve the adaptable nature of the system without penalizing resource
heavy systems. I would be happy to review patches from anyone who wishes
to undertake it.
The main fragmentation issue we've had in the past has been due to
mbuf+cluster caching, which prevented mbufs from being freed usefully in some
cases. Jeff's ongoing work on variable-sized mbufs would entirely eliminate
that problem...
I'm going to get back to this soon as infiniband gets to a useful state
for doing high performance network testing. This is only because I have
no 10gigE but do have ib and have funding to cover working on it. I hope
to have some results and activity on this front by the end of the year.
I know it has been long coming.
Thanks,
Jeff
Robert
_______________________________________________
freebsd-hackers@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-hackers
To unsubscribe, send any mail to "freebsd-hackers-unsubscr...@freebsd.org"
