Optimization: Keep some requests between GC.

[nyaxt]

Allocating buffer of bufSize is typically large enough to kick off GC, and GC flushes all pooled messages.
To avoid too much GC kicking in, let 2 messages survive between GC.

This change makes https://github.com/nyaxt/otaru faster by 30%.

[bufdev]

@tv42 I have to second this one, big time. For comparison, I wrote a "reference" fuse implementation that is effectively a pass-through to the OS file system, and I've only done extremely simple benchmarks, but using this fork makes write performance approximately 2x faster.

@nyaxt furthermore, I tested values of minPooledReq by powers of two with some basic memory profiling. With the current bazil.org/fuse implementation, 99.5.% of allocated memory (-alloc_space) is in allocMessage in these benchmarks. With this fork, it drops to about 97.5%. But if you change minPooledReq to 32, it drops to about 53%, and there's a performance benefit. The benefits drop off after 32 (ie 64 isn't really better, 128 is not).

This might indicate a more overall design discussion, but just of note:

1. Yes, please merge this :)
2. Maybe we should discuss the overall design?

I'm going to send up the dummy implementation as a separate PR.

[tv42]

Sorry for being too silent. I'm looking at very similar problems right now, and realized this probably isn't prominent enough:

Long-term plan for performance, as far as Linux is concerned, is vmsplice: #35

That gets us rid of managing per-request buffers completely.

[tv42]

9e8e652 should help here. Leaving this pull open until I've actually had time to experiment on whether it's useful even after the decreased gc churn.