First of all thank you very much to the Julia and package developers. I really enjoy this programming language!
So far I have successfully used Julia's CUDA package on my workstation containing a single CUDA device. However, I would like to use the CUDA cluster of my university, where a single node holds 6 CUDA devices. The job scheduler is slurm, which stores as many CUDA device IDs of unused GPUs in the environment variable "CUDA_VISIBLE_DEVICES" as are requested by a job. When a job (1 CPU + 1 GPU) is assigned the device ID "0" it runs fine, but with any other ID it fails with: ERROR: CuDriverError(101) in CuDevice at ~/.julia/CUDA/src/devices.jl:11 If I request all 6 CUDA devices of a node, I can run 6 copies of my Julia script (each a different CUDA device ID) within the same job in parallel without any trouble. So far I have checked that the device IDs assigned by slurm match those reported as free by "nvidia-smi". Regarding slurm configuration it looks ok to me, especially since other users have single GPU jobs running on the cluster. I have also tried using CUDA 4.1 instead of 5.0, but it didn't help. So, I wonder whether perhaps Julia's CUDA package at some point assumes a certain device ID or a range of device IDs starting always with "0" or something like that? Thanks for your help!
