Open kerim371 opened 12 months ago
The problem seems to be similar to that one #43309.
What makes you think so? The only thing in common is a segmentation fault in the same external library, but the function triggering the segmentation fault is completely different.
There's next to nothing we can do about this if you don't provide a reliable reproducer, but this is very likely an upstream bug in OpenBLAS, so there's even less we can do about it. Since you're using an Intel CPU, you might want to try MKL.jl
to use MKL instead of OpenBLAS for running BLAS operations.
@giordano hi,
As segfault is caused by the same libopenblas I thought these two cases may be connected.
There's next to nothing we can do about this if you don't provide a reliable reproducer,
I understand :( BLAS is used by third party library and it is not easy to reproduce. Also this segfault is absolutely random. Sometimes the it happen and sometimes not.
But anyway thank you for MKL.jl probably this will help.
You can also consider BLISBLAS.jl. I think Julia should drop OpenBLAS (not because of your bug), and only provide generic matmul (etc., which would have "fixed" your bug) until you opt into BLISBLAS or whatever, possibly should bundle it with.
@PallHaraldsson thank you for suggestion! I hope this will help
Hi,
I use third party library and it uses
LinearAlgebra
andBLAS
. I happen to pseudo-randomly encounter a segmentation fault:The problem seems to be similar to that one #43309.
I've tried to set
@everywhere BLAS.set_num_threads(1)
after the thirdd party library is included and sometimes this helps and sometimes not.I do the calculations on the cloud with one master node and 4 computationla nodes. Each computational node has 4 cores (Intel Ica Lake).
versioninfo()
output:I installed julia LTS 1.6.7 with
jill
and updated it withusing UpdateJulia; update_julia()
.The error is very annoying :( Computation takes much time and then sudden crush...
Appreciate any help how to solve this.