Improve performance

[crzidea]

Buffer#slice() in node is extremely slow and should avoid using it! Unfortunately, we didn't realize it when create this project. So, we should have a plan to improve the performance of this module:

• Step 1: Using BufferList(bl) instead of raw Buffer to avoid calling Buffer#slice().
  I have a patch for this module, and tested it in production. In my case, I have reduced CPU usage from 70% to 50%, and memory usage from 400m to 200m, and increased throughput from 200k/min to 400k/min.
• Step 2: Replace binary module. The module is 5 years old and have not be maintained for 5 years! kafka-node is heavily depended on binary but unfortunately binary used Buffer#slice(), a lot.

This patch complete Step 1, and I hope someone else could help me with the Step 2.

[subzero10]

My opinion is that for this type of code you should write something yourself, which you know is fully optimized. This is protocol level and will not be changed often. However, BufferList looks quite good.

[hyperlink]

Awesome @crzidea! Looking forward to testing this!

[hyperlink]

There's a definite difference in performance.

Single consumer consuming 2.1 million messages:

30 partitions
before: 154s
after: 150s

1 partition
before: 194s
after: 189s

👏🏻

[crzidea]

I will Merge this PR and start working on Step 2.

[crzidea]

@hyperlink The difference in performance would be more convincing if you test on large messages.