对在aarch64 Linux环境编译安装的CinderX补充测试

前文最后说,CinderX报错不能用,这不对,我在其github存储库上提了这个issue,alexmalyshev回复

I think that's actually just a warning that you're getting but things should be working after that?Right, this is just a logged warning that CinderX tried to enable huge pages but wasn't able to. It falls back to using normal pages.

(我认为那实际上只是一个警告,在那之后一切应该仍能正常运行?是的,这只是一个记录在案的警告,表示 CinderX 尝试启用大页但未能成功,因此它会回退到使用普通页面。)

下面是测试结果

复制代码
nbs@kylin-pc:~/par$ sudo docker start gcc142
输入密码         
gcc142
nbs@kylin-pc:~/par$ sudo docker exec -it gcc142 bash
root@kylin-pc:/# cd /par/uv314
root@kylin-pc:/par/uv314# source .venv/bin/activate

cinderx.jit.compile_after_n_calls(1)
(uv314) root@kylin-pc:/par/uv314# time python ../pe932e.py
JIT: /par/cinderx-2026.3.30.0/cinderx/Jit/code_allocator.cpp:62 -- Failed to madvise [0x7fa1ffc000, 0x7fa21fc000) with MADV_HUGEPAGE, errno=22
T(16) = 72673459417881349.time=18.62825632095337
T(16) = 72673459417881349.time=9.188195705413818

real	0m27.987s
user	0m27.694s
sys	0m0.047s
-- cinderx.jit.compile_after_n_calls(0)
(uv314) root@kylin-pc:/par/uv314# time python ../pe932e.py
JIT: /par/cinderx-2026.3.30.0/cinderx/Jit/code_allocator.cpp:62 -- Failed to madvise [0x7fa0cdc000, 0x7fa0edc000) with MADV_HUGEPAGE, errno=22
T(16) = 72673459417881349.time=11.874915599822998
T(16) = 72673459417881349.time=12.66346549987793

real	0m24.587s
user	0m24.465s
sys	0m0.012s

可见不管是cinderx.jit.compile_after_n_calls(1)还是cinderx.jit.compile_after_n_calls(0),JIT后都比原始版本提速了。关于这个问题,我也发issue提问了。alexmalyshev回复

There could be a number of reasons for this, but most likely it's because letting functions run through the interpreter once allows the CPython adaptive interpreter to emit specialized opcodes (e.g. BINARY_OP_ADD_INT instead of BINARY_OP) and CinderX can make use of that to generate faster code. Feel free to run with PYTHONJITDUMPFINALHIR=1 in your environment to compare CinderX's internal representation for each function between the two runs.

In general we don't recommend compile_after_n_calls(0) for general usage as it's not meant to be performant. It's primarily helpful to test CinderX compatibility for your code.

(出现这种情况可能有多种原因,但最有可能的是:让函数通过解释器运行一次,能使 CPython 的自适应解释器发出特化的操作码(例如 BINARY_OP_ADD_INT 而非 BINARY_OP),而 CinderX 可以利用这一点来生成更快的代码。你可以在环境中设置 PYTHONJITDUMPFINALHIR=1 来比较两次运行之间 CinderX 对每个函数的内部表示差异。

一般来说,我们不建议在日常使用中设置 compile_after_n_calls(0),因为这样做的目的并不是为了追求性能。它主要用于帮助你测试代码与 CinderX 的兼容性。)

我也用他建议的PYTHONJITDUMPFINALHIR=1参数测试了,确实两者输出的内部表示差别很大。compile_after_n_calls(0)的输出非常冗长,不像认真优化过的。

还有两个方法:cinderx.jit.force_compile(fun)和cinderx.jit.lazy_compile(fun), 分别是立即编译和延迟编译,它们俩的效果没太大区别,都和compile_after_n_calls(1)的加速比一致。

复制代码
(uv314) root@kylin-pc:/par/uv314# time python ../pe932d.py

N=2499500025000000, a=24995000, b=25000000, k=8, s=49995000
T(16) = 72673459417881349,time=14.928523063659668

real	0m14.948s
user	0m14.871s
sys	0m0.004s
(uv314) root@kylin-pc:/par/uv314# 


(uv314) root@kylin-pc:/par/uv314# time python ../pe932e.py
JIT: /par/cinderx-2026.3.30.0/cinderx/Jit/code_allocator.cpp:62 -- Failed to madvise [0x7fbbb7c000, 0x7fbbd7c000) with MADV_HUGEPAGE, errno=22
T(16) = 72673459417881349.time=14.70127558708191
cinderx.jit.force_compile(solve_T16)
T(16) = 72673459417881349.time=8.867673873901367

real	0m23.639s
user	0m23.529s
sys	0m0.012s



(uv314) root@kylin-pc:/par/uv314# time python ../pe932e.py
JIT: /par/cinderx-2026.3.30.0/cinderx/Jit/code_allocator.cpp:62 -- Failed to madvise [0x7fa909c000, 0x7fa929c000) with MADV_HUGEPAGE, errno=22
T(16) = 72673459417881349.time=14.878132820129395
cinderx.jit.lazy_compile(solve_T16)
T(16) = 72673459417881349.time=9.286654710769653

real	0m24.211s
user	0m24.080s
sys	0m0.024s
相关推荐
Decay__3 小时前
PostgreSQL 数据库全栈运维实战系列 — 合集索引(24 篇)
运维·数据库·sql·mysql·架构
戊本菌3 小时前
进程(三)runqueue 与 Linux 的 O(1) 调度
linux·运维·服务器
wuyk5554 小时前
Python办公自动化-第01讲-环境搭建与自动化认知
python·自动化·pyqt
高洁015 小时前
数字孪生与计算机视觉:让机器看懂现场
人工智能·python·机器学习·transformer·迁移学习
ComPDFKit5 小时前
RAG 回答只显示 PDF 文件名?用 Python 保留引用页码
开发语言·python·pdf·rag
光电笑映5 小时前
TCP 为什么可靠?从三次握手、滑动窗口到拥塞控制全面拆解(下)
linux·网络·网络协议·tcp
yz_aiks6 小时前
Termux 使用教程:把安卓手机变成随身 Linux(装 Python、SSH 远程、跑自动化脚本)
android·linux·python·智能手机·ssh·termux
2601_967980118 小时前
Flask Web 应用
python·flask·web开发·环境变量·虚拟环境
aifannihao8 小时前
ß 大写变 SS:用 Python 算清 Ruby2Shell 的缓冲区账本
python·安全·标准库·缓冲区溢出·字符串编码
csdn_aspnet9 小时前
在 Docker 容器中本地运行 SQL Server
运维·docker·容器·sqlserver