AMD Zen 3 與 Zen 4 上 FSRM (Fast Short REP MOV) 的效能問題

前幾天 Hacker News 上討論到的一篇:「Rust std fs slower than Python? No, it's hardware (」,原文則是在「Rust std fs slower than Python!? No, it's hardware!」。

原因是作者收到回報,提到一段 Rust 寫的 code (在文章裡面的 read_file_with_opendal(),透過 OpenDAL 去讀) 比 Python 的 code 還慢 (在文章裡面的 read_file_with_normal(),直接用 Python 的 open() 開然後讀取)。

先講最後發現問題是 Zen 3 (桌機版 5 系列的 CPU) 與 Zen 4 (桌機版 7 系列的 CPU) 這兩個架構上 REP MOV 系列的指令在某些情境下 (與 offset 有關) 有效能上的問題。

FSRM 類的指令被用在 memcpy()memmove() 類的地方,算是很常見備用到的功能,這次追蹤的問題發現在 glibc 裡面用到導致效能異常。

另外也可以查到在 Linux kernel 裡面也有用到:「Linux 5.6 To Make Use Of Intel Ice Lake's Fast Short REP MOV For Faster memmove()」,所以後續應該也會有些改善的討論...

Ubuntu 這邊的 issue ticket 開在「Terrible memcpy performance on Zen 3 when using rep movsb」這,上游的 glibc 也有對應的追蹤:「30995 – Zen 4: sub-optimal memcpy on very large copies」。

從作者私下得知的消息,因為 patch space 的大小限制,AMD 可能無法提供 CPU microcode 上的 patch,直接解決問題:

However, unverified sources suggest that a fix via amd-ucode is unlikely (at least for Zen 3) due to limited patch space. If you have more information on this matter, please reach out to me.

所以目前比較可行的作法是在 glibc 裡面使用到 FSRM 的地方針對 Zen 3 與 Zen 4 放 workaround,回到原來沒有 FSRM 的方式處理:

Our only hope is to address this issue in glibc by disabling FSRM as necessary. Progress has been made on the glibc front: x86: Improve ERMS usage on Zen3. Stay tuned for updates.

另外在追蹤問題的過程遇到不同的情境,得拿出不同的 profiling 工具出來用,所以也還蠻值得看過一次有個印象:

一開始的 timeit 算是 Python 裡面簡單的 benchmark library:

接著的比較是用 command line 的工具 hyperfine 產生出來的 (給兩個 command 讓他跑),查了一下發現在 Ubuntu 官方的 apt repository 裡面有包進去 (22.04+):

再來是用 strace 追問題,這個算是經典工具了,可以拿來看 syscall 被呼叫的時間點:

到後面出現了 perf 可以拿來看更底層的資訊,像是 CPU 內 cache 的情況:

接續提到的「hotspot ASM」應該也還是 perf 輸出的格式,不過不是那麼確定... 在「perf Examples」這邊可以看到 function 的分析:

而文章裡的則是可以看到已經到 assembly 層級了:


Ingo Molnár 提出讓 Linux Kernel 編譯速度提昇的 mega patch

Hacker News 首頁上看到「Massive ~2.3k Patch Series Would Improve Linux Build Times 50~80% & Fix "Dependency Hell"」這個,對應到 mailing list 上的信件是「* [PATCH 0000/2297] [ANNOUNCE, RFC] "Fast Kernel Headers" Tree -v1: Eliminate the Linux kernel's "Dependency Hell"」這個,看到「0000/2297」這個 prefix XDDD

他主要是想要改善 Linux Kernel 的 compile 時間 (從 project 的名稱「Fast Kernel Headers」可以看到),只是沒想到會縮短這麼多。另外一方面也順便處理了 dependency hell 的問題 (改善維護性)。

測試出來的結果相當驚人,從 231.34 +- 0.60 secs (15.5 builds/hour) 到 129.97 +- 0.51 secs (27.7 builds/hour),以編譯次數來看的話是 78% 的改善。如果以 CPU time 來看的話,從 11,474,982.05 msec cpu-clock 降到 7,100,730.37 msec cpu-clock,也是以編譯次數來算的話,有 61.6% 的改善...



When I started this project, late 2020, I expected there to be maybe 50-100 patches. I did a few crude measurements that suggested that about 20% build speed improvement could be gained by reducing header dependencies, without having a substantial runtime effect on the kernel. Seemed substantial enough to justify 50-100 commits.


But as the number of patches increased, I saw only limited performance increases. By mid-2021 I got to over 500 commits in this tree and had to throw away my second attempt (!), the first two approaches simply didn't scale, weren't maintainable and barely offered a 4% build speedup, not worth the churn of 500 patches and not worth even announcing.


With the third attempt I introduced the per_task() machinery which brought the necessary flexibility to reduce dependencies drastically, and it was a type-clean approach that improved maintainability. But even at 1,000 commits I barely got to a 10% build speed improvement. Again this was not something I felt comfortable pushing upstream, or even announcing. :-/


But the numbers were pretty clear: 20% performance gains were very much possible. So I kept developing this tree, and most of the speedups started arriving after over 1,500 commits, in the fall of 2021. I was very surprised when it went beyond 20% speedup and more, then arrived at the current 78% with my reference config. There's a clear super-linear improvement property of kernel build overhead, once the number of dependencies is reduced to the bare minimum.

這次的 patch 雖然超大包,但看起來對於 compile 時間改善非常多,應該會有不少討論... 消息還蠻新的 (台灣時間今天早上五點的信),晚點可以看一下其他大老出來回什麼...

Fork 自微軟的 Pyjion 專案的 Python 3.10 + JIT 方案

Hacker News 上看到「Pyjion – A Python JIT Compiler (」這個專案,也是一個想要透過 JIT 加速 Python 的專案:

Pyjion is a drop-in JIT Compiler for Python 3.10. It can be pip installed into a CPython 3.10 installation on Linux, Mac OS X, or Windows.

看了一下是從微軟的 Pyjion 專案 fork 出來的,原來的專案最後一次 commit 是一年前,而且專案也已經標示為 archived (read-only mode),但有留下轉移的說明,也就是上面提到的專案:

Development has moved to

可以看到大部分的效能都已經進入改善階段 (很多導入 JIT 的專案在初期時會先變慢):

跟其他的 JIT 方案相比,Pyjion 的目標是高度相容現有 Python 的程式,包括各種 extension,這點的確是在用 PyPy 這些軟體時的痛點沒錯...

看起來透過 pip 裝好後就可以直接 import 進來用,後續就會生效:

import pyjion; pyjion.enable()

另外提一下,翻 Hacker News 留言的時候翻到這個害我笑出來,有夠新 XD

zatarc 3 days ago | unvote | prev | next [–]

Pyjion requires: CPython 3.10 and .NET 6

.NET 6 Release: 19 hours ago (

... ok.

Cloudflare 也推出自己的 Speed Test 服務

Cloudflare 推出了自己的 Speed Test 服務:「Test your home network performance」。

這個服務跟 Netflix 推出的 類似,測試的是使用者端到 Netflix (或是 Cloudflare) 中間的速度,主要的目的還是公關 (PR),所以看看就好,實際上用 Speedtest 測出來會比較有參考價值,而且可以選擇不同的點測試...

不過這讓我想到之前有人測出來遠傳會對偵測使用者要使用 Speedtest 測試時開放速限的情況 (像是「遠傳吃到飽只有 Speedtest 沒限速」這篇),然後就有各種定時去打 Speedtest 觸發開放速限的方法...

目前好像只剩下這篇活著,內文提到的是 Android 上的方法,另外推文有人提到 iOS 下的方法:「[心得] 在Android破解遠X限速」,如果有遇到的可以用看看...

JavaScript 的壓縮器 esbuild

esbuild 是個 JavaScript bundler & minifier,在 GitHub 上的副標提到了重點在於速度:

An extremely fast JavaScript bundler and minifier


另外從最終的檔案大小也可以看出來,與最小的 rollup + terser 組合沒有差太多:

實際拿個 jQuery 跑看看,可以看出來壓縮的效果還行:

-rw-r--r-- 1 gslin staff  89228 Feb 19 06:03 jquery-3.4.1-esbuild.min.js
-rw-r--r-- 1 gslin staff 280364 May  2  2019 jquery-3.4.1.js
-rw-r--r-- 1 gslin staff  88145 May  2  2019 jquery-3.4.1.min.js

速度主要是透過 Golang 並且平行化運算達到的:

  • It's written in Go, a language that compiles to native code
  • Parsing, printing, and source map generation are all fully parallelized
  • Everything is done in very few passes without expensive data transformations
  • Code is written with speed in mind, and tries to avoid unnecessary allocations

不過作者有提到這個專案畢竟比較新,還沒有被時間磨練過,可能會有些 bug:

This is a hobby project that I wrote over the 2019-2020 winter break. I believe that it's relatively complete and functional. However, it's brand new code and probably has a lot of bugs. It also hasn't yet been used in production by anyone. Use at your own risk.

可以先放一陣子看看,讓一些先賢先烈把比較大的 bug 踩一踩修一修...

Amazon Aurora 支援快速複製

Amazon Aurora 宣佈支援快速複製:「Amazon Aurora Fast Database Cloning」。

對於 2TB 的資料大約五分鐘就完成了:

This means my 2TB snapshot restore job that used to take an hour is now ready in about 5 minutes – and most of that time is spent provisioning a new RDS instance.

主要是得力於後端 storage 的部份可以實做 copy-on-write 架構:

By taking advantage of Aurora’s underlying distributed storage engine you’re able to quickly and cheaply create a copy-on-write clone of your database.

可以快速複製就可以很快的驗證一些事情,像是可以直接測試 ALTER TABLE 需要的時間,或是事前演練...

用 Go 寫的 Badger

Dgraph 在推銷自家發展出來的 Badger:「Introducing Badger: A fast key-value store written natively in Go」。

標靶是 RocksDB,號稱比 RocksDB 快好幾倍:

Based on benchmarks, Badger is at least 3.5x faster than RocksDB when doing random reads. For value sizes between 128B to 16KB, data loading is 0.86x - 14x faster compared to RocksDB, with Badger gaining significant ground as value size increases. On the flip side, Badger is currently slower for range key-value iteration, but that has a lot of room for optimization.

不過我覺得有些重要的功能在 Badger 不提供,這比起來有種橘子比蘋果的感覺... 像是 RocksDB 提供了 Transaction,而 Badger 則是直接講明他們不打算支援 Transaction:

Keep it simple, stupid. No support for transactions, versioning or snapshots -- anything that can be done outside of the store should be done outside.

各家 glob 的效能...

在「Glob Matching Can Be Simple And Fast Too」這邊看到在分析 (a.*)nb 這樣的 pattern 的效能 (像是 a.*a.*a.*b 這樣的東西),第一波先測 shell,結果發現有趣的現象:

那個 csh 是怎麼了 XDDD

Looking at the source code, it doesn’t attempt to perform glob expansion itself. Instead it calls the C library implementation glob(3), which runs in linear time, at least on this Linux system. So maybe we should look at programming language implementations too.


各家實做方式不一樣 XD

然後文章裡有提到這個方式是之前蠻常見的 DoS 技巧,用很少的資源就可以吃光你的 CPU resource... 另外也提到了可能的解法,就是限制星號的數量:

At the very least, most FTP servers should probably reject glob patterns with more than, say, 3 stars.


號稱目前最快的 Terminal 軟體 (因為用 GPU 加速)

看到「Announcing Alacritty, a GPU-accelerated terminal emulator」這個用 GPU 加速 rendering 的 terminal emulator:「Alacritty」。

Alacritty is a blazing fast, GPU accelerated terminal emulator. It’s written in Rust and uses OpenGL for rendering to be the fastest terminal emulator available.

全螢幕全文字的情況下可以到 500 fps:

Alacritty’s renderer is capable of doing ~500 FPS with a large screen full of text. This is made possible by efficient OpenGL usage.

現在支援 Linux 與 macOS,不過要自己編,會比較麻煩一點:

Alacritty currently supports macOS and Linux, and Windows support is planned before the 1.0 release.

nginx 的 TCP Fast Open

在「Enabling TCP Fast Open for NGINX on CentOS 7」這邊看到 nginxTCP Fast Open (TFO,RFC 7413) 的支援早在 1.5.8 就有了,而 Linux Kernel 也是 3.7 之後就全面支援了。

TCP Fast Open 利用第一次連線後產生的 TCP cookie,在第二次連線時可以在 3-way handshake 的過程就開始傳輸,藉此大幅降低 latency。

設定方法不難,先在 kernel 設定 net.ipv4.tcp_fastopen=3,再加上 fastopen=number 就可以了,像是這樣:

listen 80 fastopen=256

不過目前 NGINX Mainline 上的版本好像沒有編進去,暫時沒辦法測...