This replaces the first-fit algorithm used in the Arena with a best-fit. According to "Dynamic Storage Allocation: A Survey and Critical Review", Wilson et. al. 1995, http://www.scs.stanford.edu/14wi-cs140/sched/readings/wilson.pdf, both startegies work well in practice.
The advantage of using best-fit is that we can switch the slow O(n) algorithm to O(log(n)) operations. Additionally, some previously O(log(n)) operations are now replaced with O(1) operations by using a hash map. The end effect is that the benchmark runs about 2.5 times faster on my machine:
old: BenchLockedPool, 5, 530, 5.25749, 0.00196938, 0.00199755, 0.00198172
new: BenchLockedPool, 5, 1300, 5.11313, 0.000781493, 0.000793314, 0.00078606
I've run all unit tests and benchmarks.
Check for unreasonable alloc size in LockedPool rather than lancing through new
Arenas until we improbably find one worthy of the quixotic request or the system
can support no more Arenas.
Add a pool for locked memory chunks, replacing LockedPageManager.
This is something I've been wanting to do for a long time. The current
approach of locking objects where they happen to be on the stack or heap
in-place causes a lot of mlock/munlock system call overhead, slowing
down any handling of keys.
Also locked memory is a limited resource on many operating systems (and
using a lot of it bogs down the system), so the previous approach of
locking every page that may contain any key information (but also other
information) is wasteful.
