lock contention

[bitfaster]

In the Read+Write scenario, the bottleneck is lock contention while attempting to enter the maintenance lock.

Breakdown of time in maintenance:

Fix:

1. Try to exit the lock early if maintenance has just run and it is possible to add to the write buffer. This reduces wait time for a queue of writes when in a lock convoy.
2. When maintenance is running on the background thread, keep track of whether the write buffer was fully drained, and if not run maintenance again. This has the effect of maintenance constantly running in the background when the buffers are populated.
3. Remove the spin wait, which is just slowing things down (throughput is measurably lower).

This is an approx 10% throughput gain for the Read+Write scenario.