Optimize COMPRESS protocol by pooling / re-using byte array/buffer

I'm currently working to minimize our application UDP bandwith usage by enabling the COMPRESS protocol in Wildfly 8.2.1.Final.

Enabling the COMPRESS protocol works fine on that regards by reducing approximatly by 7x the amount of UDP traffic which is really neat.

The thing is by enabling compression we are trading off increased cpu / memory usage for lowered network bandwith usage. In our case, we actually a notable increase of the cpu / memory usage due to increased usage of the garbage collector.

When I had a look at the COMPRESS protocol implementation we see that both Deflator and Inflator are pooled and re-used (great) but not at the byte array level (not great).

This compression pattern is very similar to GZip compression filter we can find in any recent Http spec implementors. In the Netty project for example, there's various compression protocol and/or utilities that are available and fine-tunable and they all have in common to re-use byte buffer and/or try to avoid byte arrays instantiation/copy at maximum.

Here is some notable classes of the Netty implementation:
https://github.com/netty/netty/blob/4.1/codec-http/src/main/java/io/netty/handler/codec/http/HttpContentCompressor.java
https://github.com/netty/netty/blob/4.1/codec/src/main/java/io/netty/handler/codec/compression/ZlibWrapper.java
https://github.com/netty/netty/blob/4.1/codec/src/main/java/io/netty/handler/codec/compression/ZlibCodecFactory.java
https://github.com/netty/netty/blob/4.1/codec/src/main/java/io/netty/handler/codec/compression/JZlibEncoder.java

I think re-implementing something similar in JGroups would yield great benefits to the COMPRESS protocol. Maybe also, there's something that could be picked off right from the JBoss XNIO and/or the Undertow projects?

Finally, I do not know also if we could add ergonomics at the same time to automatically determine the best number of Deflator and Inflator by keeping trace of the amount of time waited to pool one of them?

Thanks,