Performance/overhead with disabled pthread vs. pthread but single threaded

[svenpilz]

Hi fellow Emscripten users,

we are trying to make use of multi threading in an application which gets called from JS, does some longer running calculations (think of a big for loop) and returns the result back to JS. Now we observed that just enabling pthreads (we also use simd, -msimd128 -pthread) without actually using multiple threads (our dispatcher falls back to single threaded execution in that case), the runtime increases by maybe 10% (about 2sec here). Also for parts of the code that don't know about threading.

Is this expected? Is there a rather big overhead for just having pthread support activated? We don't have exception support enabled and it seems there is no difference between static/dynamic memory size.

Thanks
Sven

[kripken]

Just enabling pthreads has the effect of making various things use actual locks (like malloc, file operations, etc.). Also atomic operations are actually atomic. Both of those slowdowns happen even if you never create another thread.

A 10% slowdown is more than I'd expect, though, so it might be interesting to profile in the browser devtools and see how the profiles differ between the two builds - that diff could show interesting things.

[svenpilz]

Thanks for the tip and your time. As you thought, seems it is malloc/free and their atomics. They differ by about 2sec total time. Guess we should reduce heap allocations a bit :). The code uses a lot of dynamic Eigen matrices.

With pthread and run with a single thread in a web worker.

Self Time	Total Time	Activity
850.0 ms, 4.8 %	850.0 ms, 4.8 %	a_swap.1
687.1 ms, 3.9 %	1684.8 ms, 9.5 %	dlfree
684.0 ms. 3.9 %	684.0 ms, 3.9 %	a_cas.3
599.3 ms. 3.4 %	1622.1 ms, 9.1 %	dlmalloc

Without pthread in web worker.

Self Time	Total Time	Activity
591.9 ms, 4.0 %	591.9 ms, 4.0 %	dlfree
490.4 ms, 3.3 %	491.0 ms, 3.3 %	dlmalloc

[sbc100]

Interesting. I wonder if using tcmalloc or something like that might make sense?

[Mintyboi]

@kripken I understand this discussion has been resolved. Please let me know if I should start another topic.

I just recently enabled PThreads in my application and am currently hitting into the above performance degradation. Just wanted to check if there's anything that I can do or try out.
Is the tcmalloc approach mentioned by @sbc100 something viable?

[sbc100]

These are internal aliases for the malloc functions, you ned to add these aliases to tcmalloc, like we do for emmalloc and dlmalloc. e.g.

emscripten/system/lib/emmalloc.c

Lines 829 to 830 in a0d43fc

	extern __typeof(emmalloc_malloc) emscripten_builtin_malloc __attribute__((alias("emmalloc_malloc")));
	extern __typeof(emmalloc_malloc) __libc_malloc __attribute__((alias("emmalloc_malloc")));

[svenpilz]

@sbc100, do you know if mimalloc should in theory work under Emscripten?

I got mimalloc to compile with

  void* __libc_malloc(size_t size)                      MI_FORWARD1(mi_malloc, size)
  void* __libc_calloc(size_t count, size_t size)        MI_FORWARD2(mi_calloc, count, size)
  void* __libc_realloc(void* p, size_t size)            MI_FORWARD2(mi_realloc, p, size)
  void  __libc_free(void* p)                            MI_FORWARD0(mi_free, p)
  void* __libc_memalign(size_t alignment, size_t size)  { return mi_memalign(alignment, size); }

// Added 

  void* emscripten_builtin_malloc(size_t size)                      MI_FORWARD1(mi_malloc, size)
  void  emscripten_builtin_free(void* p)                            MI_FORWARD0(mi_free, p)
  void* emscripten_builtin_memalign(size_t alignment, size_t size)  { return mi_memalign(alignment, size); }     

but it crashes on runtime with multi-threading.

[svenpilz]

Ah, that seems be design: microsoft/mimalloc#720 (comment).

[svenpilz]

Will try tcmalloc again once I get some time next in the next sprints :). I didn't get it to export malloc/free, not sure why.

[Mintyboi]

Sorry for the late reply, I did not try out tcmalloc. Found out about rpmalloc and was giving it a go just as a sort of POC to understand these issues but had some issues integrating into my application and did not had the time to continue in a small sample app.

@mannk123 did prototyped using mspace_malloc successfully

emscripten/system/lib/dlmalloc.c

Line 5604 in 90d76db

void* mspace_malloc(mspace msp, size_t bytes) {

and found that it improves performance by 5x, from ~500ms to ~100ms for millions of malloc + free operations
I'll let him share his impl details and sample app.