Add some benchmarks comparing akka-kryo-serialization with other serialization approaches available on Akka

[romix]

It would be nice to have some benchmarks demonstrating that akka-kryo-serialization is much faster than the standard serialization provided by Akka.

[romix]

@luben @briantopping Hi guys! If your or other users of this library could contribute some benchmarks demonstrating the advantages of using it, it would be really nice of you!

[luben]

yes, I will look how to add a config with Java serialization (which is the default for akka) in our performance tests)

[luben]

PR: #72

[romix]

@luben Thanks a lot!

Could we have a benchmark that starts a set of Akka actors locally and let's them communicate (e.g. 100000 messages or something like that)? The current perf test just measures the raw serialization/deserialization performance. But the question is if the speedup is as impressive when we use it in a real-world scenario where we also have other overheads due to networking, multi-threading, etc.

[luben]

It will depend a lot on the type of messages exchanged and will vary with the workload I guess. Adapting one of the stress tests from the actor test suite should be possible with a little work. I have not looked in the details how to do it but purely local actors do not serialize the messages - we should use the multi-node testkit to spin multiple JVMs.

[luben]

hm, we can force it to always serialize the messages

[romix]

Yes, we could force to always serialize the messages, but I'm not sure if the result will be realistic enough. I'd really like to see something that would go through the OS networking stack, etc.

Basically, the aim is to show that the overall overhead of the serialization in the real-world scenarios is significant and our library is able to significantly reduce it (it may be both in terms of performance and/or amount of data transmitted).

The tests could try to exchange messages of different sizes and structure to show how the performance is changing depending on the size and structure of the message.

I agree that adapting the stress tests from the actor test suite could be a good idea and save a lot of work.

[nvollmar]

Do we still want to invest time in this?

[danischroeter]

Since performance really does matter this would be worth a bit of time. But I'm more thinking about performing a manual set of tests once or once per release...