Hi, I am very new to Julia and try to use multi-threading to make my code faster. If I do not use linear algebra, my code using “Thread.@threads” for the “for” loop speeds up as expected. However, when I use linear algebra in the loop, it just does not give any speed increase. (Indeed, it gets w…

The issue might be memory allocations in the second example, which prevent you from reaping the full benefits of multithreading

As @gdalle said, allocations will likely make multithreading performance much worse. Allocations happen when creating the 10x10 rand array. If you instead preallocate the array, then it may be a much better comparison: using Random using LinearAlgebra function test() nMax = 1e2 mMax = 1e5 …

Note that allocations also happen when you compute a * a before taking the sum. But more importantly, the code @jmair doesn’t seem thread-safe, because several threads can write into the same matrix storage. I would recommend one matrix per thread

Thanks for spotting the performance issue, which has been fixed. The code is still completely thread-safe though as the inner loop all happens on the same thread and there are no shared variables.

Oh right, I had not noticed that the matrix a was created within the threaded loop. My bad!

But in “fixing the performance issue” you turned the computation of a * a into a .* a, which is not what the initial problem stated. Perhaps using the in-place multiplication mul! would be more appropriate here?

Yes this is definitely needed. Although we need a second array since inplace mul! cannot be aliased.

Note also that in this example the temporary matrices are being created for each iteration of the nMax loop, and not one per thread. That can allocate less and be somewhat faster if the workload is divided in chunks: julia> @btime test() 380.045 ms (274 allocations: 182.22 KiB) julia> using Chun…

Might not answer the exact question, but some might find it useful anyway… My experience with multithreaded BLAS is that it should only be used on very large problems. On my Linux Ryzen 3700X, it is often better to set BLAS to single threaded if I use many matrix operation in the region of 100-1000…

Thank you so much, gdalle, jimair, lmiq, and fastwave! I am VERY sorry that my response is so slow… I tried the suggested changes one by one, but still my issue remains… My code (test4.jl) now reads the same as the final one by lmiq: using LinearAlgebra using BenchmarkTools using Random using Ch…

Multi-threading of julia-1.8.5 does not improve speed when combined with BLAS

New to Julia

gdalle April 21, 2023, 10:05pm 15

Are you sure this is unexpected? Matrix multiplication is slower when multithreading in Julia - #12 by carstenbauer

Topic		Replies	Views
Julia Threads vs BLAS threads Internals & Design	16	11178	July 26, 2018
Matrix multiplication is slower when multithreading in Julia Performance question , multithreading , linearalgebra	13	4377	January 21, 2022
Poor performance multiplying many (large) matrices multithreaded Performance question , linearalgebra	11	2580	July 13, 2020
Why doesn't multithreading help here? Performance	12	1516	August 22, 2020
Data structures for threaded computing Performance	23	3002	October 23, 2019

Multi-threading of julia-1.8.5 does not improve speed when combined with BLAS

Related topics