Explore xxh algo and benchmark

[cristianbote]

This is related and in response of emotion-js/emotion#3241

cc @romgrk

The weird part -- but kinda expected -- was that on node 18.17.0 goober is constantly second but then drops with newer versions.

~/work/github/js-benchmark (master*) $ node ./index.js ./benchmarks/hashing.js 
### node: v18.17.0 ###
CASE: murmur2_unmodified: #########################.........................   50.17% (301 ops) (3.326 ms/op)
CASE: murmur2:            ######################################............   76.17% (457 ops) (2.190 ms/op)
CASE: xxh:                ##################################################  100.00% (600 ops) (1.667 ms/op)
CASE: fnv-1a:             #####################################.............   74.83% (449 ops) (2.227 ms/op)
CASE: djb2a:              ##########################################........   85.67% (514 ops) (1.946 ms/op)
CASE: goober_unmodified:  ############################################......   89.67% (538 ops) (1.861 ms/op)

~/work/github/js-benchmark (master*) $ nvm use v18.18.0 
Now using node v18.18.0 (npm v9.8.1)

~/work/github/js-benchmark (master*) $ node ./index.js ./benchmarks/hashing.js 
### node: v18.18.0 ###
CASE: murmur2_unmodified: #######################...........................   46.71% (298 ops) (3.356 ms/op)
CASE: murmur2:            ######################################............   77.74% (496 ops) (2.016 ms/op)
CASE: xxh:                ##################################################  100.00% (638 ops) (1.567 ms/op)
CASE: fnv-1a:             ###################################...............   71.16% (454 ops) (2.205 ms/op)
CASE: djb2a:              ########################################..........   81.19% (518 ops) (1.932 ms/op)
CASE: goober_unmodified:  ######################################............   77.43% (494 ops) (2.024 ms/op)

Besides that it really is peculiar what is happening on bun that runs so slow?!

Anyway, I've copy/pasted the xxh on goober's hash perf test, and these are the results:

node v22.7.0
~/work/github/goober (explore-xxh-algo-and-benchmark*) $ npm run test-perf-hash

> [email protected] test-perf-hash
> NODE_ENV=production node benchmarks/perf-hash.cjs


Starting HASH!
▸ goober original HASH x 2,830,730 ops/sec ±0.81% (94 runs sampled)
▸ goober optimized HASH x 12,688,564 ops/sec ±3.04% (88 runs sampled)
▸ goober optimized HASH with ASM hints x 13,045,723 ops/sec ±2.04% (92 runs sampled)
▸ twind HASH x 2,617,884 ops/sec ±2.34% (93 runs sampled)
▸ xxh x 6,105,582 ops/sec ±0.49% (93 runs sampled)

Fastest is: goober optimized HASH with ASM hints

and

node v18.17.0
~/work/github/goober (explore-xxh-algo-and-benchmark*) $ npm run test-perf-hash

> [email protected] test-perf-hash
> NODE_ENV=production node benchmarks/perf-hash.cjs


Starting HASH!
▸ goober original HASH x 2,735,026 ops/sec ±3.08% (89 runs sampled)
▸ goober optimized HASH x 13,061,639 ops/sec ±0.89% (90 runs sampled)
▸ goober optimized HASH with ASM hints x 12,969,513 ops/sec ±1.12% (92 runs sampled)
▸ twind HASH x 2,653,173 ops/sec ±0.68% (93 runs sampled)
▸ xxh x 6,017,804 ops/sec ±0.73% (90 runs sampled)

Fastest is: goober optimized HASH

Looking forward to your thoughts

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
goober-rocks	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Aug 26, 2024 6:56am

[codesandbox-ci]

This pull request is automatically built and testable in CodeSandbox.

To see build info of the built libraries, click here or the icon next to each commit SHA.

[github-actions]

Size Change: 0 B

Total Size: 3.84 kB

ℹ️ View Unchanged

Filename	Size	Change
dist/goober.esm.js	1.26 kB	0 B
dist/goober.modern.js	1.26 kB	0 B
dist/goober.umd.js	1.32 kB	0 B

_{compressed-size-action}

[romgrk]

I'm exploring the disassembly of xxh on V8, it looks like it's doing a lot of validation and doesn't have confidence yet that it's indeed a Smi value despite all the existing asm.js annotations, so I went overboard and added them everywhere

It does seem to help all engines:

xxh and xxh_unmodified in this commit, I'll just keep the optimized version to continue the exploration.

---

I've also established in emotion-js/emotion#3241 (comment) that the final string formatting function can have a big impact, so goober doing a simple 'go' + number might be more efficient than xxh doing .toString(36).

---

I need to explore more before giving a better answer but I'll include the updated goober version in my benchmarks, though the collision rate makes it a bad candidate for emotion. I'll try to figure if I can find a better algo for your requirements (super low byte size).

[romgrk]

Note that you don't return anything from the goober hash functions in your tests, that might also impact the result (engines could ellicit an unused constant value with no side-effect). Also your benchmark has a single constant string as dataset instead of a proper parameter, that's another factor that could be used for optimization by the engines. It would be better to build a representative datasets of strings you see in production, and run the perf tests on that dataset.

[romgrk]

For the record, here is a confirmation that 'go' + number (decimal formatting) is very efficient on v8. No significant difference in other engines:

[cgiosy]

Hello. I slightly modified the code to make xxh32 suitable for benchmarking, and obtained the following results:

# Node (v22.6.0)

Starting HASH!
▸ twind HASH x 3,099,745 ops/sec ±0.40% (100 runs sampled)
▸ goober original HASH x 5,780,316 ops/sec ±0.44% (95 runs sampled)
▸ goober optimized HASH x 14,695,271 ops/sec ±0.67% (95 runs sampled)
▸ goober optimized HASH with ASM hints x 14,728,415 ops/sec ±0.54% (95 runs sampled)
▸ xxh32 x 17,213,817 ops/sec ±0.31% (98 runs sampled)

Fastest is: xxh32


# Bun (1.1.34)

Starting HASH!
▸ twind HASH x 8,288,343 ops/sec ±0.74% (94 runs sampled)
▸ goober original HASH x 1,672,256 ops/sec ±0.17% (98 runs sampled)
▸ goober optimized HASH x 2,075,684 ops/sec ±0.96% (93 runs sampled)
▸ goober optimized HASH with ASM hints x 10,446,495 ops/sec ±0.90% (93 runs sampled)
▸ xxh32 x 25,015,040 ops/sec ±1.87% (86 runs sampled)

Fastest is: xxh32

Actually, considering the bundle size of goober, using xxh32 is overkill. I just wrote this because there was talk about my library. :)