Cumulation optimization

[normanmaurer]

This changes try to minimize the size of the cumulation buffer to a minimum. Before we used a DynamicChannelBuffer which was never "shrinked" again. The problem with the previous implementation was that it could lead to very big cumulation buffers.

For example if you receive a buffer with 65k it will create a cumulation buffer with min. 65k + whatsleft. If you are lucky enough you will even get a cumulation buffer of 2 * 65k as the DynamicChannelBuffer will double its size if it needs more room.

The changes here try to fix it.

@trustin please review

[normanmaurer]

Seems like the speed and memory usage is really good..

See:
https://twitter.com/#!/timfox/status/194465361066860544

[trustin]

We shouldn't use a composite buffer - it's known to perform pretty bad and it will make hotspot confused because of one more ChannelBuffer implementation to consider.

[normanmaurer]

We could also just use a HeapChannelBuffer but I'm not sure this would be good in terms of memory usage and speed (byte copies)

[trustin]

I realized that your change limits the number of components in the composite buffer to 2. Because the number of components are very small, I think it might be OK.

[normanmaurer]

@trustin so ok to pull in ?

[trustin]

Yes, once buffer factory issues are fixed.

[trustin]

Ah, could we compare the performance between the composite buffer version and the heap buffer version?

[normanmaurer]

Let me try to write one but Tim's perf tests are promising with CompositeChannelBuffer

[trustin]

Yeah, I just wonder if it's gonna be even faster with the heap buffer. Could you ask Time for the numbers? :-)

[trustin]

We'll have to rewrite this part when we switch to a heap buffer.

+                // create a new buffer and copy the readable buffer into it
+                this.cumulation = newCumulationBuffer(ctx, buf.readableBytes());
+                this.cumulation.writeBytes(buf);

[normanmaurer]

Why?

[normanmaurer]

@trustin Tim tried his benchmark also with HeapChannelBuffer but he saw no speed improvements. So I think we should merge it like it is

[trustin]

I see. Thanks!