As per cache - I hardly understand affected logic, so my review wouldn't
help much here.
As per the rest changes - looks good for me. I also see garbage from NIO
and "force keys" as huge memory hotspots. The only problem is
GridCompoundFuture:
if (futs == null)
futs = new ArrayList<>();
futs.add(fut);
Things like this are proven to be anti-pattern in terms of memory
allocations, because on the last line you effectively allocate Object[10],
while usually you will have much less child futures. We could delay
allocation of array if we store the very first child future as direct
reference. And only second added future should lead to ArrayList
allocation. This should positively affect lots operations with "compound
semantics" and single cache key involved (e.g. single puts/gets).
On Mon, Feb 1, 2016 at 9:08 PM, Yakov Zhdanov <yzhda...@apache.org> wrote:
> No visible changes to throughput and latency on our common configuration,
> but allocation pressure reduced up to 20% in put-get benchmarks.
>
> --Yakov
>
> 2016-02-01 20:02 GMT+03:00 Dmitriy Setrakyan <dsetrak...@apache.org>:
>
> > Any preliminary performance numbers?
> >
> > On Mon, Feb 1, 2016 at 8:52 AM, Yakov Zhdanov <yzhda...@apache.org>
> wrote:
> >
> > > Vladimir Ozerov and Alex Goncharuk, can you please take a look at PR
> and
> > > provide comments? Other reviewers are welcome, too! =)
> > >
> > > https://github.com/apache/ignite/pull/422
> > >
> > > I did some changes to decrease allocation pressure and fixed force keys
> > > request not to be sent if rebalancing has already been successfully
> > > completed on operation's topology version.
> > >
> > > --Yakov
> > >
> >
>
