Exploration of WebAssembly multi-threading on plain merge sort
To build a native executable, run build.sh (on Windows, run it in Git Bash) or invoke your compiler directly, e.g. with sanitizers:
g++ src/main.cpp -std=c++17 -g -O0 -lpthread s -fsanitize=thread
or
g++ src/main.cpp -std=c++17 -g -O0 -lpthread s -fsanitize=address
To build WebAssembly, compile with Emscripten (I used version 1.38.16). For single-threaded version, build as:
emcc -std=c++17 -Os -DNDEBUG --llvm-lto 1 -s NO_FILESYSTEM=1 -s TOTAL_MEMORY=1073741824 -o t.js main.cpp
and for multi-threaded runs it is necessary to enable filesystem emulation (for some reason, it is required even without memory init file) and set threading related options, e.g. for 4 threads:
emcc -std=c++17 -Os --llvm-lto 1 -s USE_PTHREADS=1 -s PTHREAD_POOL_SIZE=7 -s PROXY_TO_PTHREAD=1 -DNDEBUG -s TOTAL_MEMORY=1073741824 --memory-init-file 0 -o t.js main.cpp
Then create index.html, e.g.
<!DOCTYPE html>
<html>
<title>Threads test</title>
<body>
<script src="t.js"></script>
</body>
</html>and serve it with some HTTP server, e.g. http-server that can be installed with npm:
npm i -g http-server
Then open localhost:8080 in Chrome with enabled WebAssembly thread support, as described here:
https://developers.google.com/web/updates/2018/10/wasm-threads
and look in the Console (F12).
Native code shows about 3x speed up when sorting 10,000,000 floats on 4 threads. When running 8 threads on a 4-core CPU, oversubscription constitutes itself in slight slowdown. WebAssembly running in Chrome 72 shows the expected performance with a single thread, about 1.5...2 times as slow as the native code. But, if left and right temporary arrays are allocated on the stack as std::vector<value_t> in each call to merge(), multi-threaded WebAssembly performance suffers greatly:
From the ather hand, pre-allocating temporary array ahead of time results in WebAssembly performance scaling similar to that of the native code:
Sorting on a single thread takes native code 1.55 seconds, WebAssembly 2.4...3.2 seconds, manually written JavaScript 2.62 seconds, and Kotlin/JVM (Kotlin 1.3.21 with JRE 1.8) 1.9 seconds:
