From d5e8e10947951dde5b01f53751f7fa1ce1b0fe7f Mon Sep 17 00:00:00 2001 From: Kyle Xiao Date: Tue, 6 Aug 2024 22:24:04 +0800 Subject: [PATCH] refactor: finalize reflect & README.md --- README.md | 79 ++++++++++++++++++++---------------- README_cn.md | 81 +++++++++++++++++++++---------------- frugal_jit.go | 10 +++++ internal/reflect/decoder.go | 12 +++--- options.go | 14 ------- 5 files changed, 109 insertions(+), 87 deletions(-) diff --git a/README.md b/README.md index e1e9fe5..f5e5ff3 100644 --- a/README.md +++ b/README.md @@ -2,50 +2,61 @@ English | [中文](README_cn.md) -A very fast dynamic Thrift serializer & deserializer based on just-in-time compilation. +A very fast dynamic Thrift serializer & deserializer without generating code. + +It implements a pure Go version (or reflect version) and a just-in-time(JIT) compilation version. +Since the reflect version performs better than the JIT version in most cases and it works on different cpu architectures, +so we plan to deprecate the JIT version starting with Go 1.24. + +Before Go 1.24: +* JIT version is the default serializer & deserializer for `amd64`. +* reflect version is only enabled when running on `!amd64` or `windows` +* You can enable the reflect version by specifying os env `FRUGAL_NO_JIT=1` + +Since Go 1.24: +* reflect version is the default serializer & deserializer. +* JIT code will be skipped by go build tags. + ## Features ### Code Generation Free -Traditional Thrift serializer and deserializer are based on generated code which is no longer needed since we can use JIT compilation to dynamically generate machine code. +Traditional Thrift serializer and deserializer are based on generated code which is no longer needed since we can make use of struct field tags. ### High Performance -Thanks to JIT compilation, Frugal can generate better machine code than Go language compiler. In multi-core scenarios, Frugal's performance is about 5 times higher than that of traditional serializer and deserializer. +Based on the test cases in `frugal/tests`, Frugal's performance is 1 to 4 times better than Apache Thrift (TBinaryProtocol). + +There may be variations between different test cases. Feel free to share your test cases with us. ```text -name old time/op new time/op delta -MarshalAllSize_Parallel/small-16 78.8ns ± 0% 14.9ns ± 0% -81.10% -MarshalAllSize_Parallel/medium-16 1.34µs ± 0% 0.32µs ± 0% -76.32% -MarshalAllSize_Parallel/large-16 37.7µs ± 0% 9.4µs ± 0% -75.02% -UnmarshalAllSize_Parallel/small-16 368ns ± 0% 30ns ± 0% -91.90% -UnmarshalAllSize_Parallel/medium-16 11.9µs ± 0% 0.8µs ± 0% -92.98% -UnmarshalAllSize_Parallel/large-16 233µs ± 0% 21µs ± 0% -90.99% - -name old speed new speed delta -MarshalAllSize_Parallel/small-16 7.31GB/s ± 0% 38.65GB/s ± 0% +428.84% -MarshalAllSize_Parallel/medium-16 12.9GB/s ± 0% 54.7GB/s ± 0% +322.10% -MarshalAllSize_Parallel/large-16 11.7GB/s ± 0% 46.8GB/s ± 0% +300.26% -UnmarshalAllSize_Parallel/small-16 1.56GB/s ± 0% 19.31GB/s ± 0% +1134.41% -UnmarshalAllSize_Parallel/medium-16 1.46GB/s ± 0% 20.80GB/s ± 0% +1324.55% -UnmarshalAllSize_Parallel/large-16 1.89GB/s ± 0% 20.98GB/s ± 0% +1009.73% - -name old alloc/op new alloc/op delta -MarshalAllSize_Parallel/small-16 112B ± 0% 0B -100.00% -MarshalAllSize_Parallel/medium-16 112B ± 0% 0B -100.00% -MarshalAllSize_Parallel/large-16 779B ± 0% 57B ± 0% -92.68% -UnmarshalAllSize_Parallel/small-16 1.31kB ± 0% 0.10kB ± 0% -92.76% -UnmarshalAllSize_Parallel/medium-16 448B ± 0% 3022B ± 0% +574.55% -UnmarshalAllSize_Parallel/large-16 1.13MB ± 0% 0.07MB ± 0% -93.54% - -name old allocs/op new allocs/op delta -MarshalAllSize_Parallel/small-16 1.00 ± 0% 0.00 -100.00% -MarshalAllSize_Parallel/medium-16 1.00 ± 0% 0.00 -100.00% -MarshalAllSize_Parallel/large-16 1.00 ± 0% 0.00 -100.00% -UnmarshalAllSize_Parallel/small-16 6.00 ± 0% 1.00 ± 0% -83.33% -UnmarshalAllSize_Parallel/medium-16 6.00 ± 0% 30.00 ± 0% +400.00% -UnmarshalAllSize_Parallel/large-16 4.80k ± 0% 0.76k ± 0% -84.10% +go version go1.22.4 darwin/arm64 + +goos: linux +goarch: amd64 +pkg: github.com/cloudwego/frugal/tests +cpu: Intel(R) Xeon(R) Gold 5118 CPU @ 2.30GHz + +BenchmarkAllSize_Marshal_ApacheThrift/small-4 2070745 584.0 ns/op 998.32 MB/s 112 B/op 1 allocs/op +BenchmarkAllSize_Marshal_ApacheThrift/medium-4 78729 13680 ns/op 1280.57 MB/s 112 B/op 1 allocs/op +BenchmarkAllSize_Marshal_ApacheThrift/large-4 3097 376184 ns/op 1179.75 MB/s 620 B/op 1 allocs/op +BenchmarkAllSize_Marshal_Frugal_JIT/small-4 4939591 242.1 ns/op 2407.83 MB/s 13 B/op 0 allocs/op +BenchmarkAllSize_Marshal_Frugal_JIT/medium-4 160820 7485 ns/op 2340.29 MB/s 54 B/op 0 allocs/op +BenchmarkAllSize_Marshal_Frugal_JIT/large-4 5370 214258 ns/op 2071.35 MB/s 338 B/op 0 allocs/op +BenchmarkAllSize_Marshal_Frugal_Reflect/small-4 10171197 117.3 ns/op 4970.90 MB/s 0 B/op 0 allocs/op +BenchmarkAllSize_Marshal_Frugal_Reflect/medium-4 180207 6644 ns/op 2636.73 MB/s 0 B/op 0 allocs/op +BenchmarkAllSize_Marshal_Frugal_Reflect/large-4 6312 185534 ns/op 2392.04 MB/s 0 B/op 0 allocs/op + +BenchmarkAllSize_Unmarshal_ApacheThrift/small-4 768525 1443 ns/op 403.94 MB/s 1232 B/op 5 allocs/op +BenchmarkAllSize_Unmarshal_ApacheThrift/medium-4 24463 47067 ns/op 372.19 MB/s 44816 B/op 176 allocs/op +BenchmarkAllSize_Unmarshal_ApacheThrift/large-4 1053 1155725 ns/op 384.00 MB/s 1135540 B/op 4433 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_JIT/small-4 2575767 466.3 ns/op 1250.36 MB/s 547 B/op 2 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_JIT/medium-4 62128 19333 ns/op 906.12 MB/s 19404 B/op 89 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_JIT/large-4 2328 496431 ns/op 893.99 MB/s 495906 B/op 2283 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_Reflect/small-4 2770252 437.2 ns/op 1333.60 MB/s 544 B/op 1 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_Reflect/medium-4 64232 18183 ns/op 963.45 MB/s 19945 B/op 57 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_Reflect/large-4 2325 496415 ns/op 894.02 MB/s 511876 B/op 1467 allocs/op ``` ## What can you do with Frugal ? diff --git a/README_cn.md b/README_cn.md index dc4d179..6daae84 100644 --- a/README_cn.md +++ b/README_cn.md @@ -2,50 +2,63 @@ [English](README.md) | 中文 -一款基于 JIT 编译的高性能动态 Thrift 编解码器。 +一种无需生成代码、高性能的动态 Thrift 编解码器。 + +它实现了纯 Go 版本（或reflect版本）和 即时编译（JIT）版本。 +由于在大多数情况下，reflect版本的性能优于 JIT 版本，并且它可以在不同的 CPU 架构上运行， +因此我们计划从 Go 1.24 开始废弃 JIT 版本。 + +在 Go 1.24 之前： + +* JIT 版本是 `amd64` 的默认编解码器。 +* reflect 版本仅在运行在 `!amd64` 或 `windows` 上时启用 +* 您可以通过指定操作系统环境变量 `FRUGAL_NO_JIT=1` 来启用reflect版本 + +自 Go 1.24 起： + +* reflect 版本成为默认编解码器。 +* JIT 代码将在 go build 时跳过不执行 ## 特点 ### 无需生成代码 -传统的 Thrift 编解码方式，要求用户必须要先生成编解码代码，Frugal 通过 JIT 编译技术在运行时动态生成编解码机器代码，避免了这一过程。 +传统的 Thrift 编解码方式，要求用户必须要先生成编解码代码，Frugal 通过反射 struct field tag 动态生成编解码器避免了这一过程。 ### 高性能 -基于 JIT 技术 Frugal 可以生成比 Go 语言编译器性能更好的机器代码，在多核场景下，Frugal 的性能可以达到传统编解码方式的 5 倍左右。 +基于 `frugal/tests` 的测试用例，Frugal 的性能 比 Apache Thrift (TBinaryProtocol) 好 1 到 4 倍。 + +不同的测试用例，结果可能会有些差异。欢迎给我们分享你的测试数据。 + ```text -name old time/op new time/op delta -MarshalAllSize_Parallel/small-16 78.8ns ± 0% 14.9ns ± 0% -81.10% -MarshalAllSize_Parallel/medium-16 1.34µs ± 0% 0.32µs ± 0% -76.32% -MarshalAllSize_Parallel/large-16 37.7µs ± 0% 9.4µs ± 0% -75.02% -UnmarshalAllSize_Parallel/small-16 368ns ± 0% 30ns ± 0% -91.90% -UnmarshalAllSize_Parallel/medium-16 11.9µs ± 0% 0.8µs ± 0% -92.98% -UnmarshalAllSize_Parallel/large-16 233µs ± 0% 21µs ± 0% -90.99% - -name old speed new speed delta -MarshalAllSize_Parallel/small-16 7.31GB/s ± 0% 38.65GB/s ± 0% +428.84% -MarshalAllSize_Parallel/medium-16 12.9GB/s ± 0% 54.7GB/s ± 0% +322.10% -MarshalAllSize_Parallel/large-16 11.7GB/s ± 0% 46.8GB/s ± 0% +300.26% -UnmarshalAllSize_Parallel/small-16 1.56GB/s ± 0% 19.31GB/s ± 0% +1134.41% -UnmarshalAllSize_Parallel/medium-16 1.46GB/s ± 0% 20.80GB/s ± 0% +1324.55% -UnmarshalAllSize_Parallel/large-16 1.89GB/s ± 0% 20.98GB/s ± 0% +1009.73% - -name old alloc/op new alloc/op delta -MarshalAllSize_Parallel/small-16 112B ± 0% 0B -100.00% -MarshalAllSize_Parallel/medium-16 112B ± 0% 0B -100.00% -MarshalAllSize_Parallel/large-16 779B ± 0% 57B ± 0% -92.68% -UnmarshalAllSize_Parallel/small-16 1.31kB ± 0% 0.10kB ± 0% -92.76% -UnmarshalAllSize_Parallel/medium-16 448B ± 0% 3022B ± 0% +574.55% -UnmarshalAllSize_Parallel/large-16 1.13MB ± 0% 0.07MB ± 0% -93.54% - -name old allocs/op new allocs/op delta -MarshalAllSize_Parallel/small-16 1.00 ± 0% 0.00 -100.00% -MarshalAllSize_Parallel/medium-16 1.00 ± 0% 0.00 -100.00% -MarshalAllSize_Parallel/large-16 1.00 ± 0% 0.00 -100.00% -UnmarshalAllSize_Parallel/small-16 6.00 ± 0% 1.00 ± 0% -83.33% -UnmarshalAllSize_Parallel/medium-16 6.00 ± 0% 30.00 ± 0% +400.00% -UnmarshalAllSize_Parallel/large-16 4.80k ± 0% 0.76k ± 0% -84.10% +go version go1.22.4 darwin/arm64 + +goos: linux +goarch: amd64 +pkg: github.com/cloudwego/frugal/tests +cpu: Intel(R) Xeon(R) Gold 5118 CPU @ 2.30GHz + +BenchmarkAllSize_Marshal_ApacheThrift/small-4 2070745 584.0 ns/op 998.32 MB/s 112 B/op 1 allocs/op +BenchmarkAllSize_Marshal_ApacheThrift/medium-4 78729 13680 ns/op 1280.57 MB/s 112 B/op 1 allocs/op +BenchmarkAllSize_Marshal_ApacheThrift/large-4 3097 376184 ns/op 1179.75 MB/s 620 B/op 1 allocs/op +BenchmarkAllSize_Marshal_Frugal_JIT/small-4 4939591 242.1 ns/op 2407.83 MB/s 13 B/op 0 allocs/op +BenchmarkAllSize_Marshal_Frugal_JIT/medium-4 160820 7485 ns/op 2340.29 MB/s 54 B/op 0 allocs/op +BenchmarkAllSize_Marshal_Frugal_JIT/large-4 5370 214258 ns/op 2071.35 MB/s 338 B/op 0 allocs/op +BenchmarkAllSize_Marshal_Frugal_Reflect/small-4 10171197 117.3 ns/op 4970.90 MB/s 0 B/op 0 allocs/op +BenchmarkAllSize_Marshal_Frugal_Reflect/medium-4 180207 6644 ns/op 2636.73 MB/s 0 B/op 0 allocs/op +BenchmarkAllSize_Marshal_Frugal_Reflect/large-4 6312 185534 ns/op 2392.04 MB/s 0 B/op 0 allocs/op + +BenchmarkAllSize_Unmarshal_ApacheThrift/small-4 768525 1443 ns/op 403.94 MB/s 1232 B/op 5 allocs/op +BenchmarkAllSize_Unmarshal_ApacheThrift/medium-4 24463 47067 ns/op 372.19 MB/s 44816 B/op 176 allocs/op +BenchmarkAllSize_Unmarshal_ApacheThrift/large-4 1053 1155725 ns/op 384.00 MB/s 1135540 B/op 4433 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_JIT/small-4 2575767 466.3 ns/op 1250.36 MB/s 547 B/op 2 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_JIT/medium-4 62128 19333 ns/op 906.12 MB/s 19404 B/op 89 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_JIT/large-4 2328 496431 ns/op 893.99 MB/s 495906 B/op 2283 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_Reflect/small-4 2770252 437.2 ns/op 1333.60 MB/s 544 B/op 1 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_Reflect/medium-4 64232 18183 ns/op 963.45 MB/s 19945 B/op 57 allocs/op +BenchmarkAllSize_Unmarshal_Frugal_Reflect/large-4 2325 496415 ns/op 894.02 MB/s 511876 B/op 1467 allocs/op ``` ## 用 Frugal 可以做什么？ diff --git a/frugal_jit.go b/frugal_jit.go index 416af02..6f65eae 100644 --- a/frugal_jit.go +++ b/frugal_jit.go @@ -19,7 +19,9 @@ package frugal import ( + "os" "reflect" + "strconv" "sync" "github.com/cloudwego/frugal/internal/jit/decoder" @@ -30,6 +32,14 @@ import ( "github.com/cloudwego/gopkg/protocol/thrift" ) +var nojit bool + +func init() { + if v, err := strconv.ParseBool(os.Getenv("FRUGAL_NO_JIT")); err == nil { + nojit = v + } +} + func jitEncodedSize(val interface{}) int { return encoder.EncodedSize(val) } diff --git a/internal/reflect/decoder.go b/internal/reflect/decoder.go index 343e388..94f392f 100644 --- a/internal/reflect/decoder.go +++ b/internal/reflect/decoder.go @@ -225,7 +225,6 @@ func (d *tDecoder) decodeType(t *tType, b []byte, p unsafe.Pointer, maxdepth int // tmpv = decode(b) // map[tmpk] = tmpv tmp := t.MapTmpVarsPool.Get().(*tmpMapVars) - defer t.MapTmpVarsPool.Put(tmp) k := tmp.k v := tmp.v kp := tmp.kp @@ -247,6 +246,8 @@ func (d *tDecoder) decodeType(t *tType, b []byte, p unsafe.Pointer, maxdepth int sliceV = d.Malloc(l*vt.V.Size, vt.V.Align, vt.V.MallocAbiType) } + var n int + var err error i := 6 for j := 0; j < l; j++ { p = kp @@ -260,8 +261,8 @@ func (d *tDecoder) decodeType(t *tType, b []byte, p unsafe.Pointer, maxdepth int if kt.FixedSize > 0 { i += decodeFixedSizeTypes(kt.T, b[i:], p) } else { - if n, err := d.decodeType(kt, b[i:], p, maxdepth-1); err != nil { - return i, err + if n, err = d.decodeType(kt, b[i:], p, maxdepth-1); err != nil { + break } else { i += n } @@ -277,14 +278,15 @@ func (d *tDecoder) decodeType(t *tType, b []byte, p unsafe.Pointer, maxdepth int if vt.FixedSize > 0 { i += decodeFixedSizeTypes(vt.T, b[i:], p) } else { - if n, err := d.decodeType(vt, b[i:], p, maxdepth-1); err != nil { - return i, err + if n, err = d.decodeType(vt, b[i:], p, maxdepth-1); err != nil { + break } else { i += n } } m.SetMapIndex(k, v) } + t.MapTmpVarsPool.Put(tmp) // no defer, it may be in hot path return i, nil case tLIST, tSET: // NOTE: for tSET, it may be map in the future // list header diff --git a/options.go b/options.go index 4e49b76..304c871 100644 --- a/options.go +++ b/options.go @@ -18,24 +18,10 @@ package frugal import ( "fmt" - "os" - "runtime" - "strconv" "github.com/cloudwego/frugal/internal/opts" ) -// Frugal supports `reflect` mode which faster than jit version in many scenarios, -// and performs more stable in concurrency, also works in different CPU architectures. -// It will become a default option soon. -var nojit = runtime.GOARCH != "amd64" // always nojit=true under non-amd64 env - -func init() { - if v, err := strconv.ParseBool(os.Getenv("FRUGAL_NO_JIT")); err == nil { - nojit = v - } -} - const ( _MinILSize = 1024 )