performance reduced in 6.x

[tyan-boot]

After upgrade to version 6.0.1, it seems the latency increased, here is my benchmark result using criterion:

Maps/dash-map/read/v5   time:   [165.64 ns 171.17 ns 176.44 ns]
                        change: [-4.5163% +1.9547% +8.3860%] (p = 0.55 > 0.05)
                        No change in performance detected.
Maps/dash-map/read/v6   time:   [342.38 ns 354.98 ns 366.45 ns]
                        change: [-7.5138% +2.9050% +15.397%] (p = 0.61 > 0.05)
                        No change in performance detected.

This my code used to benchmark:

use criterion::{BatchSize, Criterion, criterion_group, criterion_main};
use dashmap::DashMap;
use dashmap6::{DashMap as DashMap6};

type MyID = [u8; 5];

fn bench_maps(c: &mut Criterion) {
    let mut group = c.benchmark_group("Maps");
    group.bench_function("dash-map/read/v5", |b| {
        b.iter_batched(
            || {
                let maps = DashMap::new();
                maps.insert(MyID::from([1, 2, 3, 4, 5]), MyID::default());
                maps
            },
            |maps| {
                criterion::black_box(maps.get(&MyID::from([1, 2, 3, 4, 5])));
            },
            BatchSize::SmallInput,
        )
    });
    group.bench_function("dash-map/read/v6", |b| {
        b.iter_batched(
            || {
                let maps = DashMap6::new();
                maps.insert(MyID::from([1, 2, 3, 4, 5]), MyID::default());
                maps
            },
            |maps| {
                criterion::black_box(maps.get(&MyID::from([1, 2, 3, 4, 5])));
            },
            BatchSize::SmallInput,
        )
    });
}

criterion_group!(benches, bench_maps);
criterion_main!(benches);

Cargo.toml

[dependencies]
dashmap = { version = "5.5.3", features = ["inline"] }
dashmap6 = { version = "6", package = "dashmap", features = ["inline"] }


[dev-dependencies]
criterion = "0.5.1"

[[bench]]
name = "map"
path = "benches/map.rs"
harness = false

Environment

Tested on Apple M1, M2, both running macOS
Also test on AMD 5600G which runs Windows 11

Rust version output by rustc -vV

rustc 1.79.0 (129f3b996 2024-06-10)
binary: rustc
commit-hash: 129f3b9964af4d4a709d1383930ade12dfe7c081
commit-date: 2024-06-10
host: aarch64-apple-darwin
release: 1.79.0
LLVM version: 18.1.7

[tyan-boot]

And bench result on Windows11 AMD 5600G

Maps/dash-map/read/v5   time:   [212.48 ns 236.06 ns 262.47 ns]
                        change: [-18.107% -9.0717% +0.9671%] (p = 0.08 > 0.05)
                        No change in performance detected.

Maps/dash-map/read/v6   time:   [273.19 ns 301.81 ns 331.36 ns]
                        change: [-7.7231% +2.7725% +15.169%] (p = 0.65 > 0.05)

[cxw620]

Update: notice commit 92d64ba78c2e0d418a7dc16abb89872f7e8f7301, PR #303 cause such regression.

Here's the test result after I revert the commit:

Maps/dash-map/read/v5   time:   [349.53 ns 361.70 ns 372.76 ns]
                        change: [-14.103% -1.2692% +12.890%] (p = 0.85 > 0.05)
                        No change in performance detected.
Maps/dash-map/read/v6   time:   [166.15 ns 189.84 ns 210.08 ns]
                        change: [-16.376% -2.6358% +12.527%] (p = 0.74 > 0.05)
                        No change in performance detected.

---

Notice great pref regression, too. About 2x slower, similar to tyan's result on Apple platform.

Maps/dash-map/read/v5   time:   [361.47 ns 375.36 ns 387.71 ns]
                        change: [-11.272% +1.3747% +16.931%] (p = 0.85 > 0.05)
                        No change in performance detected.
Maps/dash-map/read/v6   time:   [663.62 ns 713.58 ns 756.31 ns]
                        change: [-21.620% -4.4423% +16.276%] (p = 0.65 > 0.05)
                        No change in performance detected.

Env: Debian 12, R7-8845HS

rustc 1.81.0-nightly (24d2ac0b5 2024-07-15)
binary: rustc
commit-hash: 24d2ac0b56fcbde13d827745f66e73efb1e17156
commit-date: 2024-07-15
host: x86_64-unknown-linux-gnu
release: 1.81.0-nightly
LLVM version: 18.1.7

Profile:

opt-level = 3
debug = false
lto = true
codegen-units = 1
split-debuginfo = "packed"
strip = "debuginfo"
panic = "unwind"

[esemeniuc]

@xacrimon any thoughts?

[cxw620]

Well, according to the result of conc-map-bench (also created by xacrimon) , 6.0.1 actually outperforms 5.5.3, though slightly.

I believe this is the reason why xacrimon retains the commit 92d64ba78c2e0d418a7dc16abb89872f7e8f7301, as dashmap's primary goal is to replace rwlock + hashmap.