Hey!
So I think we have no extreme naysayers to Optional. So let me try to sum
up what we have achieved so:
- In macroscale benchmark based on REST interface using Optionals didn't
lower the performance.
- +1 for using it in public APIs, especially for those using functional
style.
- Creating lots of Optional instances might add some pressure on GC, so
we need to be careful when using them in hot code paths. In such cases it
is required to run a micro scale benchamark to make sure the performance
didn't drop. The microbenchmark should also be followed by macro scale
benchamrk - PerfJobAck. Also, keep an eye on Eden space in such cases.
If you agree with me, and there are no hard evidence that using Optional
degrade performance significantly, I would like to issue a pull request and
put those findings into contributing guide [1].
Thanks,
Sebastian
[1]
On Mon, May 22, 2017 at 6:36 PM Galder Zamarreño <galder(a)redhat.com> wrote:
I think Sanne's right here, any differences in such large scale
test are
hard to decipher.
Also, as mentioned in a previous email, my view on its usage is same as
Sanne's:
* Definitely in APIs/SPIs.
* Be gentle with it internals.
Cheers,
--
Galder Zamarreño
Infinispan, Red Hat
> On 18 May 2017, at 14:35, Sanne Grinovero <sanne(a)infinispan.org> wrote:
>
> Hi Sebastian,
>
> sorry but I think you've been wasting time, I hope it was fun :) This is
not the right methodology to "settle" the matter (unless you want Radim's
eyes to get bloody..).
>
> Any change in such a complex system will only affect the performance
metrics if you're actually addressing the dominant bottleneck. In some
cases it might be CPU, like if your system is at 90%+ CPU then it's likely
that reviewing the code to use less CPU would be beneficial; but even that
can be counter-productive, for example if you're having contention caused
by optimistic locking and you fail to address that while making something
else "faster" the performance loss on the optimistic lock might become
asymptotic.
>
> A good reason to avoid excessive usage of Optional (and *excessive*
doesn't mean a couple dozen in a millions lines of code..) is to not run
out of eden space, especially for all the code running in interpreted mode.
>
> In your case you've been benchmarking a hugely complex beast, not least
over REST! When running the REST Server I doubt that allocation in eden is
your main problem. You just happened to have a couple Optionals on your
path; sure performance changed but there's no enough data in this way to
figure out what exactly happened:
> - did it change at all or was it just because of a lucky optimisation?
(The JIT will always optimise stuff differently even when re-running the
same code)
> - did the overall picture improve because this code became much *less*
slower?
>
> The real complexity in benchmarking is to accurately understand why it
changed; this should also tell you why it didn't change more, or less..
>
> To be fair I actually agree that it's very likely that C2 can make any
performance penalty disappear.. that's totally possible, although it's
unlikely to be faster than just reading the field (assuming we don't need
to do branching because of null-checks but C2 can optimise that as well).
> Still this requires the code to be optimised by JIT first, so it won't
prevent us from creating a gazillion of instances if we abuse its usage
irresponsibly. Fighting internal NPEs is a matter of writing better code;
I'm not against some "Optional" being strategically placed but I believe
it's much nicer for most internal code to just avoid null, use "final",
and
initialize things aggressively.
>
> Sure use Optional where it makes sense, probably most on APIs and SPIs,
but please don't go overboard with it in internals. That's all I said in
the original debate.
>
> In case you want to benchmark the impact of Optional make a JMH based
microbenchmark - that's interesting to see what C2 is capable of - but even
so that's not going to tell you much on the impact it would have to patch
thousands of code all around Infinispan. And it will need some peer review
before it can tell you anything at all ;)
>
> It's actually a very challenging topic, as we produce libraries meant
for "anyone to use" and don't get to set the hardware specification
requirements it's hard to predict if we should optimise the system for
this/that resource consumption. Some people will have plenty of CPU and
have problems with us needing too much memory, some others will have the
opposite.. the real challenge is in making internals "elastic" to such
factors and adaptable without making it too hard to tune.
>
> Thanks,
> Sanne
>
>
>
> On 18 May 2017 at 12:30, Sebastian Laskawiec <slaskawi(a)redhat.com>
wrote:
> Hey!
>
> In our past we had a couple of discussions about whether we should or
should not use Optionals [1][2]. The main argument against it was
performance.
>
> On one hand we risk additional object allocation (the Optional itself)
and wrong inlining decisions taken by C2 compiler [3]. On the other hand we
all probably "feel" that both of those things shouldn't be a problem and
should be optimized by C2. Another argument was the Optional's doesn't give
us anything but as I checked, we introduced nearly 80 NullPointerException
bugs in two years [4]. So we might consider Optional as a way of fighting
those things. The final argument that I've seen was about lack of higher
order functions which is simply not true since we have #map, #filter and
#flatmap functions. You can do pretty amazing things with this.
>
> I decided to check the performance when refactoring REST interface. I
created a PR with Optionals [5], ran performance tests, removed all
Optionals and reran tests. You will be surprised by the results [6]:
>
> Test case
> With Optionals [%] Without Optionals
> Run 1 Run 2 Avg Run 1 Run 2 Avg
> Non-TX reads 10 threads
> Throughput 32.54 32.87 32.71 31.74 34.04 32.89
> Response time -24.12 -24.63 -24.38 -24.37 -25.69 -25.03
> Non-TX reads 100 threads
> Throughput 6.48 -12.79 -3.16 -7.06 -6.14 -6.60
> Response time -6.15 14.93 4.39 7.88 6.49 7.19
> Non-TX writes 10 threads
> Throughput 9.21 7.60 8.41 4.66 7.15 5.91
> Response time -8.92 -7.11 -8.02 -5.29 -6.93 -6.11
> Non-TX writes 100 threads
> Throughput 2.53 1.65 2.09 -1.16 4.67 1.76
> Response time -2.13 -1.79 -1.96 0.91 -4.67 -1.88
>
> I also created JMH + Flight Recorder tests and again, the results showed
no evidence of slow down caused by Optionals [7].
>
> Now please take those results with a grain of salt since they tend to
drift by a factor of +/-5% (sometimes even more). But it's very clear the
performance results are very similar if not the same.
>
> Having those numbers at hand, do we want to have Optionals in Infinispan
codebase or not? And if not, let's state it very clearly (and write it into
contributing guide), it's because we don't like them. Not because of
performance.
>
> Thanks,
> Sebastian
>
> [1]
http://lists.jboss.org/pipermail/infinispan-dev/2017-March/017370.html
> [2]
http://lists.jboss.org/pipermail/infinispan-dev/2016-August/016796.html
> [3]
http://vanillajava.blogspot.ro/2015/01/java-lambdas-and-low-latency.html
> [4]
https://issues.jboss.org/issues/?jql=project%20%3D%20ISPN%20AND%20issuety...
> [5]
https://github.com/infinispan/infinispan/pull/5094
> [6]
https://docs.google.com/a/redhat.com/spreadsheets/d/1oep6Was0FfvHdqBCwpCF...
> [7]
https://github.com/infinispan/infinispan/pull/5094#issuecomment-296970673
> --
> SEBASTIAN ŁASKAWIEC
> INFINISPAN DEVELOPER
> Red Hat EMEA
>
>
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev(a)lists.jboss.org
>
https://lists.jboss.org/mailman/listinfo/infinispan-dev
>
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev(a)lists.jboss.org
>
https://lists.jboss.org/mailman/listinfo/infinispan-dev
_______________________________________________
infinispan-dev mailing list
infinispan-dev(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev