Reduce garbage during GelfMessage encoding and writing

[mp911de]

GelfMessage increases GC pressure when encoding and sending messages which originate in several sources. This ticket is to improve GC pressure by:

• Omitting byte[] allocation
• Reusing ByteBuffer
• Revisit JSON serialization to operate directly on ByteBuffer
• Investigate on BigDecimal removal

Once these steps are performed, more complex refactorings (String to byte conversion, ByteBuffer pooling and caching/reusing of Deflater) can be tackled.

[mp911de]

Before:

Benchmark                                          Mode  Cnt      Score      Error  Units
GelfMessageAssemblerPerf.createMessage             avgt    5    137,836 ±   48,410  ns/op
GelfMessageAssemblerPerf.createMessageToJSON       avgt    5   1870,298 ± 2874,998  ns/op
GelfMessageAssemblerPerf.createMessageToTCPBuffer  avgt    5   1870,715 ± 2702,764  ns/op
GelfMessageAssemblerPerf.createMessageToUDPBuffer  avgt    5  18395,612 ± 1540,763  ns/op

After:

Benchmark                                                      Mode  Cnt      Score      Error  Units
GelfMessageAssemblerPerf.createMessage                         avgt    5     86,524 ±   21,244  ns/op
GelfMessageAssemblerPerf.createMessageToJSON                   avgt    5   1572,840 ± 2775,274  ns/op
GelfMessageAssemblerPerf.createMessageToJSONPooledDirect       avgt    5    413,643 ±   46,868  ns/op
GelfMessageAssemblerPerf.createMessageToJSONPooledHeap         avgt    5    500,867 ±  131,854  ns/op
GelfMessageAssemblerPerf.createMessageToTCPBuffer              avgt    5   1403,524 ± 2989,169  ns/op
GelfMessageAssemblerPerf.createMessageToTCPBufferPooledHeap    avgt    5    489,318 ±   55,937  ns/op
GelfMessageAssemblerPerf.createMessageToTCPPooledDirect        avgt    5    411,686 ±  117,468  ns/op
GelfMessageAssemblerPerf.createMessageToUDPBuffer              avgt    5  18001,048 ± 3506,724  ns/op
GelfMessageAssemblerPerf.createMessageToUDPBufferPooledDirect  avgt    5   6908,736 ±  749,133  ns/op
GelfMessageAssemblerPerf.createMessageToUDPBufferPooledHeap    avgt    5   6645,161 ±  256,350  ns/op

TCP and UDP senders use serialization with pooled buffers.