*: use FnvHashMap again

[overvenus]

Benchmark at https://github.com/overvenus/benches-rust

Base on the benchmark, fnv hash map outperforms ordermap and fnv ordermap.

test collections::bench_fnv_map_little                 ... bench:         165 ns/iter (+/- 7)
test collections::bench_ordermap_little                ... bench:         387 ns/iter (+/- 9)
test collections::bench_fnv_ordermap_little            ... bench:         337 ns/iter (+/- 11)

test collections::bench_fnv_map_large                  ... bench:      20,211 ns/iter (+/- 495)
test collections::bench_ordermap_large                 ... bench:      42,226 ns/iter (+/- 589)
test collections::bench_fnv_ordermap_large             ... bench:      37,202 ns/iter (+/- 860)

test collections::bench_fnv_map_query_16               ... bench:          15 ns/iter (+/- 0)
test collections::bench_fnv_map_query_32               ... bench:          17 ns/iter (+/- 0)
test collections::bench_fnv_map_query_64               ... bench:          15 ns/iter (+/- 0)
test collections::bench_fnv_map_query_128              ... bench:          15 ns/iter (+/- 0)
test collections::bench_fnv_map_query_512              ... bench:          16 ns/iter (+/- 0)

test collections::bench_ordermap_query_16              ... bench:          23 ns/iter (+/- 0)
test collections::bench_ordermap_query_32              ... bench:          23 ns/iter (+/- 0)
test collections::bench_ordermap_query_64              ... bench:          24 ns/iter (+/- 0)
test collections::bench_ordermap_query_128             ... bench:          24 ns/iter (+/- 0)
test collections::bench_ordermap_query_512             ... bench:          25 ns/iter (+/- 0)

test collections::bench_fnv_ordermap_query_16          ... bench:          18 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_query_32          ... bench:          18 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_query_64          ... bench:          18 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_query_128         ... bench:          19 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_query_512         ... bench:          20 ns/iter (+/- 0)

test collections::bench_fnv_map_delete_insert_16       ... bench:          42 ns/iter (+/- 1)
test collections::bench_fnv_map_delete_insert_32       ... bench:          34 ns/iter (+/- 0)
test collections::bench_fnv_map_delete_insert_64       ... bench:          34 ns/iter (+/- 0)
test collections::bench_fnv_map_delete_insert_128      ... bench:          44 ns/iter (+/- 1)
test collections::bench_fnv_map_delete_insert_512      ... bench:          38 ns/iter (+/- 1)

test collections::bench_ordermap_delete_insert_16      ... bench:          61 ns/iter (+/- 0)
test collections::bench_ordermap_delete_insert_32      ... bench:          64 ns/iter (+/- 2)
test collections::bench_ordermap_delete_insert_64      ... bench:          64 ns/iter (+/- 1)
test collections::bench_ordermap_delete_insert_128     ... bench:          63 ns/iter (+/- 0)
test collections::bench_ordermap_delete_insert_512     ... bench:          67 ns/iter (+/- 1)

test collections::bench_fnv_ordermap_delete_insert_16  ... bench:          55 ns/iter (+/- 2)
test collections::bench_fnv_ordermap_delete_insert_32  ... bench:          54 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_delete_insert_64  ... bench:          56 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_delete_insert_128 ... bench:          56 ns/iter (+/- 1)
test collections::bench_fnv_ordermap_delete_insert_512 ... bench:          61 ns/iter (+/- 2)

[BusyJay]

Why OrderMap is prefered now? The result seems conflict with the previous benchmark, so which is correct?

[overvenus]

Actually, fnv hashmap is prefered now :-P
It's up to compilation strategy. The previous benchmark compiled with CARGO_INCREMENTAL=1 that makes ordermap faster than fnv hashmap. Since TiKV release builds without incremental, this PR is correct!

[overvenus]

FYI, rust has an issue (rust-lang/rust#39773) about CARGO_INCREMENTAL=1 performance.

[siddontang]

any performance improved for TiKV?

[overvenus]

FnvHashMap slightly improves performance.

RawKV puts on 1 TiKV.

	OrderMap	FnvHashMap
TPS (per sec. avg(2min))	65180.04	65326.89
raftstore CPU	~86.7%	~85.3%
apply worker CPU	~97.8%	~97.9%

[arthurprs]

Probably influenced by this PR as well rust-lang/rust#40561

[overvenus]

Thanks for pointing out! @arthurprs

Your implementation has been adopted since nightly-2017-04-07. I did benchmarks for both nightly-2017-04-07 and nightly-2017-04-06 (the old). 04-07 is faster!

Left: 04-06
Right: 04-07

test collections::bench_fnv_map_delete_insert_128            39 ns/iter (+/- 0)          36 ns/iter (+/- 0)
test collections::bench_fnv_map_delete_insert_16             37 ns/iter (+/- 1)          40 ns/iter (+/- 0)
test collections::bench_fnv_map_delete_insert_32             35 ns/iter (+/- 0)          30 ns/iter (+/- 0)
test collections::bench_fnv_map_delete_insert_512            41 ns/iter (+/- 0)          36 ns/iter (+/- 0)
test collections::bench_fnv_map_delete_insert_64             44 ns/iter (+/- 0)          35 ns/iter (+/- 0)
test collections::bench_fnv_map_large                    19,216 ns/iter (+/- 179)    17,992 ns/iter (+/- 164)
test collections::bench_fnv_map_little                      144 ns/iter (+/- 2)         149 ns/iter (+/- 3)
test collections::bench_fnv_map_query_128                    18 ns/iter (+/- 0)          17 ns/iter (+/- 0)
test collections::bench_fnv_map_query_16                     17 ns/iter (+/- 0)          15 ns/iter (+/- 0)
test collections::bench_fnv_map_query_32                     17 ns/iter (+/- 0)          15 ns/iter (+/- 0)
test collections::bench_fnv_map_query_512                    22 ns/iter (+/- 0)          19 ns/iter (+/- 0)
test collections::bench_fnv_map_query_64                     17 ns/iter (+/- 0)          15 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_delete_insert_128       48 ns/iter (+/- 0)          45 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_delete_insert_16        45 ns/iter (+/- 0)          54 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_delete_insert_32        46 ns/iter (+/- 0)          43 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_delete_insert_512       47 ns/iter (+/- 1)          49 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_delete_insert_64        43 ns/iter (+/- 0)          44 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_large               26,292 ns/iter (+/- 265)    24,857 ns/iter (+/- 334)
test collections::bench_fnv_ordermap_little                 215 ns/iter (+/- 4)         215 ns/iter (+/- 6)
test collections::bench_fnv_ordermap_query_128               19 ns/iter (+/- 0)          20 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_query_16                18 ns/iter (+/- 0)          18 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_query_32                20 ns/iter (+/- 0)          19 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_query_512               20 ns/iter (+/- 0)          20 ns/iter (+/- 0)
test collections::bench_fnv_ordermap_query_64                20 ns/iter (+/- 0)          19 ns/iter (+/- 0)
test collections::bench_ordermap_delete_insert_128           48 ns/iter (+/- 0)          48 ns/iter (+/- 0)
test collections::bench_ordermap_delete_insert_16            53 ns/iter (+/- 0)          53 ns/iter (+/- 1)
test collections::bench_ordermap_delete_insert_32            48 ns/iter (+/- 1)          49 ns/iter (+/- 0)
test collections::bench_ordermap_delete_insert_512           52 ns/iter (+/- 0)          52 ns/iter (+/- 0)
test collections::bench_ordermap_delete_insert_64            47 ns/iter (+/- 0)          48 ns/iter (+/- 0)
test collections::bench_ordermap_large                   30,014 ns/iter (+/- 336)    29,758 ns/iter (+/- 314)
test collections::bench_ordermap_little                     242 ns/iter (+/- 6)         240 ns/iter (+/- 2)
test collections::bench_ordermap_query_128                   24 ns/iter (+/- 0)          24 ns/iter (+/- 0)
test collections::bench_ordermap_query_16                    24 ns/iter (+/- 0)          24 ns/iter (+/- 0)
test collections::bench_ordermap_query_32                    24 ns/iter (+/- 0)          24 ns/iter (+/- 0)
test collections::bench_ordermap_query_512                   26 ns/iter (+/- 0)          25 ns/iter (+/- 0)
test collections::bench_ordermap_query_64                    25 ns/iter (+/- 0)          25 ns/iter (+/- 0)

[siddontang]

I think we can reuse FNV again

What is your opinion, @BusyJay?

[ngaut]

LGTM

[BusyJay]

LGTM