Benchmarks

[JustinDrake]

Hi there. We're looking to use this library for Ethereum 2.0. Do you have rough performance benchmarks, e.g. for pairings?

[meyer9]

I'll upload some real benchmarks when I can set it up. Full pairings (miller loop and final exponetiation) is at about 8,000,000 ns/op vs. Zcash at about 3,000,000 ns/op. There's definitely room for improvement still.

[meyer9]

goos: linux
goarch: amd64
pkg: github.com/phoreproject/bls
BenchmarkBLSAggregateSignature-8   	  500000	      2307 ns/op
BenchmarkBLSSign-8                 	    1000	   1334142 ns/op
BenchmarkBLSVerify-8               	     100	  17212764 ns/op
BenchmarkFQ12MulAssign-8           	  100000	     16863 ns/op (6950 ns/op)
BenchmarkFQ12SquareAssign-8        	  100000	     12647 ns/op (4900 ns/op)
BenchmarkFQ12InverseAssign-8       	   20000	     60300 ns/op (25000 ns/op)
BenchmarkFQ2AddAssign-8            	20000000	      84.6 ns/op (35 ns/op)
BenchmarkFQ2SubAssign-8            	20000000	      75.2 ns/op (32 ns/op)
BenchmarkFQ2MulAssign-8            	 3000000	       562 ns/op (278 ns/op)
BenchmarkFQ2SquareAssign-8         	 3000000	       492 ns/op (220 ns/op)
BenchmarkFQ2InverseAssign-8        	   50000	     29321 ns/op (13610 ns/op)
BenchmarkFQ6AddAssign-8            	10000000	       245 ns/op
BenchmarkFQ6SubAssign-8            	10000000	       194 ns/op
BenchmarkFQ6MulAssign-8            	  300000	      4536 ns/op
BenchmarkFQ6SquareAssign-8         	  300000	      3939 ns/op
BenchmarkFQ6InverseAssign-8        	   50000	     37269 ns/op
BenchmarkFQAddAssign-8             	50000000	      35.8 ns/op (19 ns/op)
BenchmarkFQSubAssign-8             	50000000	      30.4 ns/op (14 ns/op)
BenchmarkFQMulAssign-8             	20000000	       117 ns/op (72 ns/op)
BenchmarkFQMul2-8                  	50000000	      25.4 ns/op
BenchmarkFQSquare-8                	10000000	       148 ns/op (66 ns/op)
BenchmarkFQInverse-8               	   50000	     32540 ns/op (13300 ns/op)
BenchmarkFQNegate-8                	30000000	      46.0 ns/op (11 ns/op)
BenchmarkFQSqrt-8                  	   20000	     81789 ns/op (40000 ns/op)
BenchmarkG1MulAssign-8             	    2000	    656593 ns/op (332700 ns/op)
BenchmarkG1AddAssign-8             	  500000	      2380 ns/op (1357 ns/op)
BenchmarkG1AddAssignMixed-8        	 1000000	      1878 ns/op (1000 ns/op)
BenchmarkG2MulAssign-8             	     500	   2725396 ns/op (1073000 ns/op)
BenchmarkG2AddAssign-8             	  100000	     10131 ns/op (4504 ns/op)
BenchmarkG2AddAssignMixed-8        	  200000	      7659 ns/op (3184 ns/op)
BenchmarkG2Prepare-8               	    3000	    575940 ns/op (240000 ns/op)
BenchmarkMillerLoop-8              	    1000	   1668151 ns/op (677000 ns/op)
BenchmarkFinalExponentiation-8     	     200	   6179057 ns/op (1800000 ns/op)
BenchmarkPairing-8                 	     200	   8413597 ns/op (2700000 ns/op)
BenchmarkMACWithCarry-8            	1000000000	      2.33 ns/op
BenchmarkSubWithCarry-8            	1000000000	      2.54 ns/op
BenchmarkAddWithCarry-8            	1000000000	      2.72 ns/op
PASS
ok  	github.com/phoreproject/bls	83.917s

I've also attached Zcash benchmarks for their BLS library.

[meyer9]

I bet this could be sped up even more if we switched compiling to use gcc-go. The problem with that is that the gc assembly is incompatible with gcc-go.

[meyer9]

Benchmarks using math/bits:

goos: linux
goarch: amd64
pkg: github.com/phoreproject/bls
BenchmarkBLSAggregateSignature-8   	  500000	      2960 ns/op
BenchmarkBLSSign-8                 	    1000	   1419106 ns/op
BenchmarkBLSVerify-8               	     100	  20543869 ns/op
BenchmarkFQ12MulAssign-8           	  100000	     20071 ns/op
BenchmarkFQ12SquareAssign-8        	  100000	     14215 ns/op
BenchmarkFQ12InverseAssign-8       	   20000	     65837 ns/op
BenchmarkFQ2AddAssign-8            	20000000	       100.0 ns/op
BenchmarkFQ2SubAssign-8            	20000000	        74.9 ns/op
BenchmarkFQ2MulAssign-8            	 2000000	       687 ns/op
BenchmarkFQ2SquareAssign-8         	 2000000	       687 ns/op
BenchmarkFQ2InverseAssign-8        	   30000	     37184 ns/op
BenchmarkFQ6AddAssign-8            	 5000000	       291 ns/op
BenchmarkFQ6SubAssign-8            	10000000	       250 ns/op
BenchmarkFQ6MulAssign-8            	  200000	      6658 ns/op
BenchmarkFQ6SquareAssign-8         	  300000	      4747 ns/op
BenchmarkFQ6InverseAssign-8        	   30000	     43820 ns/op
BenchmarkFQAddAssign-8             	30000000	        41.6 ns/op
BenchmarkFQSubAssign-8             	50000000	        35.4 ns/op
BenchmarkFQMulAssign-8             	10000000	       149 ns/op
BenchmarkFQMul2-8                  	50000000	        25.1 ns/op
BenchmarkFQSquare-8                	10000000	       120 ns/op
BenchmarkFQInverse-8               	   50000	     33650 ns/op
BenchmarkFQNegate-8                	30000000	        47.6 ns/op
BenchmarkFQSqrt-8                  	   20000	     84710 ns/op
BenchmarkG1MulAssign-8             	    2000	    756880 ns/op
BenchmarkG1AddAssign-8             	  500000	      2774 ns/op
BenchmarkG1AddAssignMixed-8        	 1000000	      2156 ns/op
BenchmarkG2MulAssign-8             	     500	   3005618 ns/op
BenchmarkG2AddAssign-8             	  100000	     12370 ns/op
BenchmarkG2AddAssignMixed-8        	  200000	      9254 ns/op
BenchmarkG2Prepare-8               	    2000	    713139 ns/op
BenchmarkMillerLoop-8              	    1000	   2074986 ns/op
BenchmarkFinalExponentiation-8     	     200	   7257563 ns/op
BenchmarkPairing-8                 	     100	  10425579 ns/op
BenchmarkMACWithCarry-8            	2000000000	         0.27 ns/op
BenchmarkSubWithCarry-8            	2000000000	         0.56 ns/op
BenchmarkAddWithCarry-8            	2000000000	         0.30 ns/op
PASS
ok  	github.com/phoreproject/bls	79.790s

[JustinDrake]

Thanks @meyer9. Do you want to be part of the next BLS12-381 standardisation call? (See drafts and minutes at https://github.com/pairingwg/bls_standard.)

[meyer9]

Yes. That would be great. My email is julianmeyer2000 [at] google email provider.

[kilic]

There is another prime field implementation in go which also uses avo generator.
https://github.com/kilic/fp

Benchmarked on 2.7 ghz i5 machine

384/Addition             8.95 ns/op
384/Multiplication    92.3 ns/op