Julia processes on a NUMA dual processor machine




I have a program that spawns a number of julia processes and runs some code on them. The code uses @async in a way that is similar to the sample pmap code in the julia documentation. The problem that I am seeing is that half the processes complete in the expected amount of time, but the other half take more than 5x the time. I am on a server machine with dual processors (each processor has 8 cores), NUMA enabled, hyperthreading disabled, and running Windows 10 (Hyper V host, no VMs running).

Are there any gotchas to using Julia and spawning Julia processes in a dual processor NUMA setting?




Ritchie! Welcome! Perhaps it’s best that you provide some actual code - a Minimum Working Example would help others help you.
Otherwise, it’s hard to know general gotchas.
Can you provide your Julia Version as well?
Also, consider using BenchmarkTools.jl for useful stats.


Thanks Miguel,

Turns out this is not Julia-related at all. It’s a hardware issue!