Made papaya ~9% faster using parking_lot

[ppmpreetham]

Made papaya ~9% faster single-thread reads & ~85% lower p99 tail latency in incremental mode by replacing Mutex with parking_lot

Latency Bench

Variant	Before (p99 values)	After (p99 values)
papaya (incremental)	4 ms 4 ms, 5 ms, 9 ms, 19 ms, 14 ms, 5 ms, 7 ms, 4 ms	4 ms 1 ms, 0 ms, 0 ms, 2 ms, 12 ms, 0 ms, 5 ms, 0 ms
papaya (blocking)	4 ms (single) 67 ms, 185 ms, 189 ms, 192 ms, 186 ms, 185 ms, 185 ms, 185 ms	2 ms (single) 32 ms, 86 ms, 87 ms, 86 ms, 90 ms, 86 ms, 87 ms, 86 ms

Single-threaded Read Bench

Map	Before	After	Relative Change
papaya	[161.52 µs 164.23 µs 167.35 µs]	[149.58 µs 150.64 µs 151.87 µs]	−7.7% to −10.0%
std	[93.431 µs 95.598 µs 97.980 µs]	[81.907 µs 82.839 µs 83.976 µs]	−13.2% to −18.5%
dashmap	[195.79 µs 199.30 µs 203.44 µs]	[184.78 µs 186.77 µs 189.01 µs]	−3.6% to −7.8%

BEFORE:

Running benches\latency.rs (target\release\deps\latency-febcbe066c63fc01.exe)
=== papaya (incremental) ===
p99 insert: 4ms
p99 concurrent insert: 4ms
p99 concurrent insert: 5ms
p99 concurrent insert: 9ms
p99 concurrent insert: 19ms
p99 concurrent insert: 14ms
p99 concurrent insert: 5ms
p99 concurrent insert: 7ms
p99 concurrent insert: 4ms
=== papaya (blocking) ===
p99 insert: 4ms
p99 concurrent insert: 67ms
p99 concurrent insert: 185ms
p99 concurrent insert: 189ms
p99 concurrent insert: 192ms
p99 concurrent insert: 186ms
p99 concurrent insert: 185ms
p99 concurrent insert: 185ms
p99 concurrent insert: 185ms
=== dashmap ===
p99 insert: 3ms
p99 concurrent insert: 3ms
p99 concurrent insert: 3ms
p99 concurrent insert: 3ms
p99 concurrent insert: 3ms
p99 concurrent insert: 3ms
p99 concurrent insert: 3ms
p99 concurrent insert: 3ms
p99 concurrent insert: 3ms
Running benches\single_thread.rs (target\release\deps\single_thread-c96b88cabe8f6003.exe)
Gnuplot not found, using plotters backend
read/papaya time: [161.52 µs 164.23 µs 167.35 µs]
Found 10 outliers among 100 measurements (10.00%)
5 (5.00%) high mild
5 (5.00%) high severe
read/std time: [93.431 µs 95.598 µs 97.980 µs]
Found 8 outliers among 100 measurements (8.00%)
7 (7.00%) high mild
1 (1.00%) high severe
read/dashmap time: [195.79 µs 199.30 µs 203.44 µs]
Found 10 outliers among 100 measurements (10.00%)
1 (1.00%) high mild
9 (9.00%) high severe

AFTER:

=== papaya (incremental) ===
p99 insert: 4ms
p99 concurrent insert: 1ms
p99 concurrent insert: 0ms
p99 concurrent insert: 0ms
p99 concurrent insert: 2ms
p99 concurrent insert: 12ms
p99 concurrent insert: 0ms
p99 concurrent insert: 5ms
p99 concurrent insert: 0ms
=== papaya (blocking) ===
p99 insert: 2ms
p99 concurrent insert: 32ms
p99 concurrent insert: 86ms
p99 concurrent insert: 87ms
p99 concurrent insert: 86ms
p99 concurrent insert: 90ms
p99 concurrent insert: 86ms
p99 concurrent insert: 87ms
p99 concurrent insert: 86ms
=== dashmap ===
p99 insert: 2ms
p99 concurrent insert: 2ms
p99 concurrent insert: 2ms
p99 concurrent insert: 2ms
p99 concurrent insert: 2ms
p99 concurrent insert: 2ms
p99 concurrent insert: 2ms
p99 concurrent insert: 2ms
p99 concurrent insert: 2ms
Running benches\single_thread.rs (target\release\deps\single_thread-c96b88cabe8f6003.exe)
Gnuplot not found, using plotters backend
read/papaya time: [149.58 µs 150.64 µs 151.87 µs]
change: [-10.011% -8.8298% -7.7247%] (p = 0.00 < 0.05)
Performance has improved.
Found 7 outliers among 100 measurements (7.00%)
4 (4.00%) high mild
3 (3.00%) high severe
read/std time: [81.907 µs 82.839 µs 83.976 µs]
change: [-18.498% -15.797% -13.173%] (p = 0.00 < 0.05)
Performance has improved.
Found 13 outliers among 100 measurements (13.00%)
8 (8.00%) high mild
5 (5.00%) high severe
read/dashmap time: [184.78 µs 186.77 µs 189.01 µs]
change: [-7.7925% -5.6063% -3.5561%] (p = 0.00 < 0.05)
Performance has improved.
Found 9 outliers among 100 measurements (9.00%)
7 (7.00%) high mild
2 (2.00%) high severe

[ppmpreetham]

The previous source code has a lot of unused functions, consider removing it and run it

[ibraheemdev]

I haven't looked at this closely, but I'm a little confused about your before and after numbers. Why are numbers other than papaya also changing?

I'm not sure if the primary impact of your change is caused by the allocation lock or the parker lock, but I would be interested in removing the parker entirely and using parking_lot_core::{park, unpark} instead, considering we're already adding the dependency.

[ibraheemdev]

The CI warnings should be gone after you rebase on top #90.

[ppmpreetham]

I'm confused about the timings for the 2nd run to be less too, I'll try testing it 100s of times and averaging them out.

Also, if time permits, I'd also try to work on parking using parking_lot_core::{park, unpark}.

[ppmpreetham]

Before vs After Parker:

the parker change does not materially affect the single-threaded read benchmark. For the latency bench, medians are mostly unchanged, and and the insert-max means are noisy because one scheduler pause can dominate a run. The only papaya incremental concurrent average moved slightly faster after the change.

When I ran:

cargo bench --bench single_thread -- --sample-size 200 --warm-up-time 2 --measurement-time 5 --quiet
cargo bench --bench latency

single_thread

Benchmark	Before mean	After mean	Change
read/papaya	153.78 us	153.43 us	-0.23%
read/std	80.136 us	81.044 us	+1.13%
read/dashmap	176.29 us	175.80 us	-0.28%

latency

Benchmark	Metric	Before mean	After mean	Change	Before median	After median
papaya incremental	reported insert max	2.00 ms	5.40 ms	+170.0%	2.00 ms	2.00 ms
papaya incremental	reported concurrent thread max	3.11 ms	2.88 ms	-7.6%	2.00 ms	1.00 ms
papaya blocking	reported insert max	2.00 ms	4.30 ms	+115.0%	2.00 ms	2.00 ms
papaya blocking	reported concurrent thread max	76.70 ms	79.19 ms	+3.2%	82.00 ms	82.00 ms
dashmap	reported insert max	2.10 ms	2.50 ms	+19.0%	2.00 ms	2.50 ms
dashmap	reported concurrent thread max	2.05 ms	2.21 ms	+7.9%	2.00 ms	2.00 ms