Reduce contention in all ByteBufferPool implementations

[lorban]

Target Jetty version(s)
9.4.x, 10.0.x, 11.0.x

Enhancement Description

All ByteBufferPool implementations use ByteBufferPool.Buckets to track ByteBuffers. The current implementation of Bucket stores buffers in a ConcurrentLinkedDeque, and calls poll() to acquire a pooled buffer and offerFirst() to release a buffer.

This means that all the contention is done at the head of the queue for no particular reason. An easy way to reduce that contention would be to release buffers at the tail of the queue by calling offer() instead of offerFirst(), this way all the acquisition contention would happen at the head while all the releasing contention would happen at the tail, basically halving the contention.

Replacing the ConcurrentLinkedDeque with a ConcurrentLinkedQueue could also improve performance an extra bit too.