MSMs slower than doing individual scalar multiplications about curves HOT 5 CLOSED

It's all benchmarking artifacts afaik: We'd seriously fucked up benchmarks, the arkworks benchmarks were not perfect either, and separately criterion was making it look like arkworks bechmarks reported the same problem. I think @swasilyev fixed the arkworks benchmarks. I hopefully convinced him arkworks needs a smaller MSM1024 bechmark for other reasons, which should avoid criterion issues, even on people's laptops. Also @achimcc will fix our messed up benchmarks.

from curves.

I'll close this for now since @swasilyev pointed out that criterion does not handle slow benchmarks well: bheisler/criterion.rs#320

We'll sort this out internally and reopen if necessary. It maybe worth having some smaller MSM benchmarks too.

from curves.

Individual muls:

Arithmetic for Bls12_381::G1/Scalar Multiplication
time: [191.29 µs 194.11 µs 199.25 µs]

Arithmetic for Bls12_381::G2/Scalar Multiplication
time: [692.69 µs 708.57 µs 725.30 µs]

1024-MSM:

MSM for Bls12_381::G1 time: [24.212 ms 24.980 ms 25.818 ms]

~8x faster than individual

MSM for Bls12_381::G2 time: [72.097 ms 72.759 ms 73.431 ms]

~10x faster

from curves.

I'll close this for now since @swasilyev pointed out that criterion does not handle slow benchmarks well:

@swasilyev actually didn't mean anything like that. What he meant is that you should respect other's time and double check before posting your exciting findings

from curves.

This is surprising; do you think this is because of the multi-threading overhead? Do you have an idea of the cost if you turn off parallel?

from curves.