Small performance regression in x/crypto XChaCha20 implementation
zp opened this issue · 2 comments
When building gocryptfs from current master source 9958b63, there is a small? (might vary cpu to cpu) performance regression with XChaCha20-Poly1305-Go
speed benchmark with 2.4.0
gocryptfs v2.4.0; go-fuse [vendored]; 2023-10-12 go1.21.3 linux/amd64
cpu: Intel(R) Core(TM) i3-2100 CPU @ 3.10GHz; no AES acceleration
AES-GCM-256-OpenSSL 228.65 MB/s (selected in auto mode)
AES-GCM-256-Go 61.51 MB/s
AES-SIV-512-Go 46.20 MB/s
XChaCha20-Poly1305-OpenSSL 485.02 MB/s
XChaCha20-Poly1305-Go 728.97 MB/s (selected in auto mode)
speed benchmark with 9958b63
gocryptfs v2.4.0-15-g9958b63; go-fuse [vendored]; 2024-03-03 go1.22.0 linux/amd64
cpu: Intel(R) Core(TM) i3-2100 CPU @ 3.10GHz; no AES acceleration
AES-GCM-256-OpenSSL 232.79 MB/s (selected in auto mode)
AES-GCM-256-Go 63.22 MB/s
AES-SIV-512-Go 48.48 MB/s
XChaCha20-Poly1305-OpenSSL 490.74 MB/s
XChaCha20-Poly1305-Go 616.01 MB/s (selected in auto mode)
Diagnosing this it appears that x/crypto has locked certain cpu acceleration/optimization behind proper flags, specifically this commit, as it was previously just always enabled (even if it wasn't actually supported by the cpu)
Go seems to build with GOAMD64=v1 by default, for what I assume is maximum compatibility
Building with GOAMD64=v2 re-enables the optimizations bringing it back to 2.4.0 performance
speed benchmark with 9958b63 and GOAMD64=v2
gocryptfs v2.4.0-15-g9958b63; go-fuse [vendored]; 2024-03-03 go1.22.0 linux/amd64
cpu: Intel(R) Core(TM) i3-2100 CPU @ 3.10GHz; no AES acceleration
AES-GCM-256-OpenSSL 232.84 MB/s (selected in auto mode)
AES-GCM-256-Go 62.21 MB/s
AES-SIV-512-Go 49.58 MB/s
XChaCha20-Poly1305-OpenSSL 487.09 MB/s
XChaCha20-Poly1305-Go 727.68 MB/s (selected in auto mode)
I propose to just build gocryptfs going forward with GOAMD64=v2 (at least for the static amd64 build on github releases), possibly providing a default v1 build for legacy support, as according to this chart v2 extensions date back to 2008.
GOAMD64=v2 sounds good. Care to add it build.bash and make a PR?
Reported to Go upstream as golang/go#67240