Pool timers to avoid allocations

[nussjustin]

Requests are somewhat allocation heavy. This should help a bit. Also has a small positive effect on requests performance (more for the old request style).

Benchstat (base 9147cfa):

name                              old time/op    new time/op    delta
Request-8                           43.9µs ± 5%    43.4µs ± 1%     ~     (p=0.690 n=5+5)
OldRequest-8                        60.0µs ± 3%    57.5µs ± 2%   -4.30%  (p=0.008 n=5+5)

name                              old alloc/op   new alloc/op   delta
Request-8                             885B ± 0%      693B ± 0%  -21.71%  (p=0.008 n=5+5)
OldRequest-8                        2.63kB ± 0%    2.44kB ± 0%   -7.27%  (p=0.000 n=5+4)

name                              old allocs/op  new allocs/op  delta
Request-8                             17.0 ± 0%      14.0 ± 0%  -17.65%  (p=0.008 n=5+5)
OldRequest-8                          55.0 ± 0%      52.0 ± 0%   -5.45%  (p=0.008 n=5+5)

Other benchmarks show no changes.

[coveralls]

Coverage decreased (-0.01%) to 94.376% when pulling 6d0118b on nussjustin:timer-pool into 9147cfa on nats-io:master.

[nussjustin]

Test failure is unrelated and also happens on master (though only 2 times in 20 local runs, 10 with and 10 without -race).

[coveralls]

Coverage increased (+0.1%) to 94.492% when pulling 4330ddd on nussjustin:timer-pool into 9147cfa on nats-io:master.

[nussjustin]

Since this works with sync.Pool testing is somewhat tricky. The test would need to disable GC (and reenable afterwards) and set GOMAXPROCS to 1 (and reset it afterwards) to prevent the pool from being emptied and to avoid rescheduling onto another P.

If you want I can push a test.

[tylertreat]

lgtm. Would be nice to have some basic tests around timerPool, though I agree it's a bit tricky to test with the GC behavior.

[nussjustin]

Pushed a new commit with a test that should work.

[coveralls]

Coverage increased (+0.1%) to 94.492% when pulling 6a1a909 on nussjustin:timer-pool into 9147cfa on nats-io:master.

[nussjustin]

Ah, okay. sync.Pool will randomly drop values when compiling with -race to prevent assuming it's behaviour. I changed the test to just check the behaviour of the timer instead.

[coveralls]

Coverage increased (+0.1%) to 94.492% when pulling 1630db7 on nussjustin:timer-pool into 9147cfa on nats-io:master.

[tylertreat]

+1