(Local)IntHistogram

[breezewish]

trafficstars

We need to implement a int version for Histogram, like Counter and Gauge, since atomic float implementation (loop + lock cmpxchg) is much slower than native atomic integer instruction (lock add).

For example, according to benchmarks:

test bench_counter_no_labels                          ... bench:          14 ns/iter (+/- 0)
test bench_int_counter_no_labels                      ... bench:           7 ns/iter (+/- 1)

When there is no concurrent write, float atomic implemented via CAS is 1x slower than int atomic.

test bench_counter_no_labels_concurrent_write         ... bench:       5,263 ns/iter (+/- 1,204)
test bench_int_counter_no_labels_concurrent_write     ... bench:         892 ns/iter (+/- 100)

When there is concurrent write, float atomic is 5x slower than int atomic.

[siddontang]

Mostly, we use the histogram to record operation duration in TiKV, using Int can help us, but we need to update Grafana.

[breezewish]

Yes, currently we use seconds and we need to update to milliseconds in order to use IntHistogram.

[hdost]

Is there a block of example code which could be used to benchmark this?

[mxinden]

@hdost today histograms don't support anything other than AtomicF64, thus there is no benchmark comparing AtomicF64 with e.g. AtomicU64. Benchmark running AtomicF64 can be found here.

[hdost]

Yes indeed I wasn't sure if the bench_counter_no_labels_concurrent_write would appear in the bench tests.

[hdost]

I've started on a rather large draft here: #388 before going too much further down the rabbit hole I wouldn't mind a once over. It doesn't compile and I am well aware of hat but I am curious if this path is worth pursing