bench test with golang std ed25519

[giskook]

hi,
I want to use the project to my go project.
1. I wrappered the project with rust lib and generate a C header.
2. make a project to test the perfermance.

I got the result.

goos: darwin
goarch: arm64
pkg: github.com/giskook/ed25519_bench
BenchmarkStdEd25519NewKey-10    	   61866	     17676 ns/op
BenchmarkNewKey-10              	   95047	     12670 ns/op
BenchmarkStdEd25519Sign-10      	   51584	     23155 ns/op
BenchmarkSign-10                	   70915	     16794 ns/op
BenchmarkStdEd25519Verify-10    	   23914	     50102 ns/op
BenchmarkVerify-10              	   32502	     36692 ns/op
PASS
ok  	github.com/giskook/ed25519_bench	9.527s

and

goos: linux
goarch: amd64
pkg: github.com/giskook/ed25519_bench
cpu: Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz
BenchmarkStdEd25519NewKey-16    	   58395	     20548 ns/op
BenchmarkNewKey-16              	   66391	     18012 ns/op
BenchmarkStdEd25519Sign-16      	   47431	     25199 ns/op
BenchmarkSign-16                	   51754	     23096 ns/op
BenchmarkStdEd25519Verify-16    	   19858	     60420 ns/op
BenchmarkVerify-16              	   22867	     52159 ns/op
PASS
ok  	github.com/giskook/ed25519_bench	9.211s

It seems the perfermance is not outstanding. Am I missing something?

my test project: https://github.com/giskook/ed25519_bench

[tarcieri]

CGo has high overhead compared to the time it takes to actually perform the cryptographic operations.

See the “Why not cgo” section of this blog post:

https://words.filippo.io/rustgo/

[giskook]

I had add a empty CGO call to my test project.It shows the cgo call only has lower overhead.you can review my project.

the empty CGO call takes ~25ns on my laptap, ~50ns on my aws machine. the real work takes about 12611ns.

the cgo call only takes about 0.2% time costs.

the result:

goos: darwin
goarch: arm64
pkg: github.com/giskook/ed25519_bench
BenchmarkStdEd25519NewKey-10           	   62973	     17480 ns/op
BenchmarkNewKey-10                     	   94894	     12611 ns/op
BenchmarkStdEd25519Sign-10             	   51811	     23087 ns/op
BenchmarkSign-10                       	   70196	     16753 ns/op
BenchmarkStdEd25519Verify-10           	   23860	     50249 ns/op
BenchmarkVerify-10                     	   32156	     37192 ns/op
BenchmarkJustTestForCGoEmptyCall-10    	45658698	        24.94 ns/op
PASS
ok  	github.com/giskook/ed25519_bench	11.610s

and

goos: linux
goarch: amd64
pkg: github.com/giskook/ed25519_bench
cpu: Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz
BenchmarkStdEd25519NewKey-16           	   58423	     20544 ns/op
BenchmarkNewKey-16                     	   66846	     18314 ns/op
BenchmarkStdEd25519Sign-16             	   47661	     25183 ns/op
BenchmarkSign-16                       	   51940	     23089 ns/op
BenchmarkStdEd25519Verify-16           	   19798	     60634 ns/op
BenchmarkVerify-16                     	   22994	     51869 ns/op
BenchmarkJustTestForCGoEmptyCall-16    	23624349	        50.83 ns/op
PASS
ok  	github.com/giskook/ed25519_bench	10.498s

[giskook]

hi, I run a the test between go crypto/ed25519(without cgo) vs ed25519-dalek:

test Method:
for go(version 1.18.2), go to the go src code, run the bench:

go test -bench .

for ed25519-dalek(1.0.1) run:

cargo bench --features="batch"

the result:

case 1

env:

go version	chip	OS	mem	Total CPU numbers
1.18.2	m1	macOS	32G	10

reuslt:

	sign	verify
go crypto/ed25519	22941 ns	50217 ns
ed25519-dalek	12868 ns	31581 ns

ed25519-dalek's sign performance improvement: 43.91%;

ed25519-dalek's verify performance improvement: 37.11%;

case 2

env:

go version	chip	OS	mem	Total CPU Numbers
1.18.2	Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz	Ubuntu	64G	16

reusult:

	sign	verify
go crypto/ed25519	25200 ns	60571 ns
ed25519-dalek	18480 ns	46782 ns

ed25519-dalek's sign performance improvement: 26.67%;

ed25519-dalek's verify performance improvement: 22.77%;