remote_cache/digest: add benchmark for sha256-simd #4547
Conversation
Provide a setup to compare minio/sha256-simd (Apache 2.0 license) performance vs the Go standard library "crypto/sha256". The `sha256-simd` library comes with 2 modes: - without server, automatically detect CPU features - with server, require Avx512 CPU features The ARM64 support is not tested. Running the benchmark against out remote executor yields ``` ==================== Test output for //server/remote_cache/digest:simd_bench_test: goos: linux goarch: amd64 cpu: Intel(R) Xeon(R) CPU @ 3.10GHz BenchmarkSIMDDigestCompute/without_SIMD/1-30 255240 5042 ns/op BenchmarkSIMDDigestCompute/with_SIMD_no_server/1-30 234526 5190 ns/op BenchmarkSIMDDigestCompute/with_SIMD_with_server/1-30 388 36804140 ns/op BenchmarkSIMDDigestCompute/without_SIMD/10-30 10000 118668 ns/op BenchmarkSIMDDigestCompute/with_SIMD_no_server/10-30 26204 64872 ns/op BenchmarkSIMDDigestCompute/with_SIMD_with_server/10-30 100 62445228 ns/op BenchmarkSIMDDigestCompute/without_SIMD/100-30 10000 193471 ns/op BenchmarkSIMDDigestCompute/with_SIMD_no_server/100-30 20247 135334 ns/op BenchmarkSIMDDigestCompute/with_SIMD_with_server/100-30 100 64685802 ns/op BenchmarkSIMDDigestCompute/without_SIMD/1000-30 14314 188163 ns/op BenchmarkSIMDDigestCompute/with_SIMD_no_server/1000-30 10000 176901 ns/op BenchmarkSIMDDigestCompute/with_SIMD_with_server/1000-30 100 212289431 ns/op BenchmarkSIMDDigestCompute/without_SIMD/10000-30 9067 658089 ns/op BenchmarkSIMDDigestCompute/with_SIMD_no_server/10000-30 10000 721403 ns/op BenchmarkSIMDDigestCompute/with_SIMD_with_server/10000-30 100 234613900 ns/op BenchmarkSIMDDigestCompute/without_SIMD/100000-30 2685 1577976 ns/op BenchmarkSIMDDigestCompute/with_SIMD_no_server/100000-30 1924 1079974 ns/op BenchmarkSIMDDigestCompute/with_SIMD_with_server/100000-30 100 146595705 ns/op BenchmarkSIMDDigestCompute/without_SIMD/1000000-30 312 9117083 ns/op BenchmarkSIMDDigestCompute/with_SIMD_no_server/1000000-30 298 13086220 ns/op BenchmarkSIMDDigestCompute/with_SIMD_with_server/1000000-30 56 211401036 ns/op PASS ================================================================================ ```
| func hasherWithServer() hash.Hash { | ||
| server := sha256simd.NewAvx512Server() | ||
| return sha256simd.NewAvx512(server) | ||
| } |
There was a problem hiding this comment.
in real usage, would we reuse this server across requests? wonder if the server should be declared as a top-level var instead of creating a new server on every iteration
There was a problem hiding this comment.
https://github.com/minio/sha256-simd/blob/master/README.md#support-for-avx512
Due to this different way of scheduling, we decided to use an explicit method to instantiate the AVX512 version. > Essentially one or more AVX512 processing servers (Avx512Server) have to be created whereby each server can hash over 3 GB/s on a single core. An hash.Hash object (Avx512Digest) is then instantiated using one of these servers and used in the regular fashion:
I think the expectation here is to create 1 server for each core? there are not a lot of examples 🤔
Since the doc mentioned speed up for cases Overall, the constraint of (1) a bigger size message, (2) server - CPU core 1-1 mapping, and (3) message padding for alignment make it quite unattractive. to our use case. Gona close this for now. |
|
After digging into this a bit more, it seems like the CPU we have on GCP, at least for our executor, do not include Intel's SHA extension And the minio/sha256-simd code has this clause https://github.com/minio/sha256-simd/blob/6096f891a77bfe490cbea7a424c821b5fdb92849/cpuid_other.go#L27 So when we use The AVX512 implementation is mostly targeted toward hashing bigger files/messages and thus is not suitable for our use case for now. The ARM64 implementation could be attractive for ARM64 executors (Linux / MacOS) down the line, but my benchmark on M1 laptop does not show a big speed-up. Pushed my latest local setup to the branch so future me / other folks could replicate the experiment. |
Provide a setup to compare minio/sha256-simd (Apache 2.0 license)
performance vs the Go standard library "crypto/sha256".
The
sha256-simdlibrary comes with 2 modes:The ARM64 support is not tested.
Running the benchmark against out remote executor yields
Related issues: N/A