Simpler MEX Multi-Threading w/ a Persistent Thread Pool

版本 1.1.2 (12.9 KB) 作者: oreoman
Further speed up your MEX files by making it easier to add in support for multi-threading and eliminate multi-threading overhead.
0.0
(0)
72.0 次下载
更新时间 2021/7/5
在 GitHub 上查看许可证
I can't believe this actually worked.
For smaller problems that can't be vectorized, properly setting up multi-threading within a MEX file can create enough overhead that the single-threaded solution is often faster. Using a thread pool eliminates that overhead. The pool is created once, and then can be re-used in multiple calls to your MEX function. The thread pool is then closed out and cleared from memory by typing 'clear all' or 'clear mex' in the command line (or closing MATLAB).
Depending on the problem, you can potentially see a speedup of up to 'maxNumCompThreads', though that's probably not likely. In my very basic testing, I'm seeing ~2x speedup (with four threads) over many iterations of a small problem. For the real-world problem I created this for, I'm seeing a ~60% speedup compared to standard MEX multi-threading, saving me a couple of days in total simulation time.
In addition, adding multi-threading to a MEX file is now much easier, as you only need to learn how to use one function, AddThreadPoolJob. You then just have to remember to close your MEX file with SynchronizeThreads so that you don't get awesome race conditions.
See the included madd_threadpool.c example and compare to madd.c that uses regular multi-threading to figure out what's going on. There's also more documentation in threadpool.h. To see how single- thread vs. multi-thread vs. thread pooled vs. vectorized code compares, compile things and check out threadpool_timings.m.
NOTE: properly vectorized MATLAB code will always be faster than this. Only use this stuff if you can't vectorize your code but it still lends itself to straightforward parallelization, e.g. integrating recurrence relations in multiple dimensions.

引用格式

oreoman (2024). Simpler MEX Multi-Threading w/ a Persistent Thread Pool (https://github.com/michael-nix/MATLAB-MEX-Threadpool), GitHub. 检索时间: .

MATLAB 版本兼容性
创建方式 R2019a
与 R2018a 及更高版本兼容
平台兼容性
Windows macOS Linux
类别
在 Help Center 和 MATLAB Answers 中查找有关 Write C Functions Callable from MATLAB (MEX Files) 的更多信息

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

无法下载基于 GitHub 默认分支的版本

版本 已发布 发行说明
1.1.2

Simple bug fix in madd_threadpool.c
1.1.1

Connected to GitHub...
1.1.0

Added a few helper functions to make my life easier, and make it possible to use the same thread pool among multiple MEX files. Not properly documented 'cause I'm lazy.
1.0.0

要查看或报告此来自 GitHub 的附加功能中的问题，请访问其 GitHub 仓库。
要查看或报告此来自 GitHub 的附加功能中的问题，请访问其 GitHub 仓库。