Using constraint to request specific GPUs
Gilbreth has heterogeneous hardware comprising of Nvidia V100, A100, A10, and A30 GPUs in separate sub-clusters. You can run sfeatures to check the specifications of different Gilbreth nodes:
NODELIST CPUS MEMORY AVAIL_FEATURES GRES
gilbreth-b[000-015] 24 190000 B,b,A30,a30 gpu:3
gilbreth-c[000-002] 20 760000 C,c,V100,v100 gpu:4
gilbreth-d[000-007] 16 190000 D,d,A30,a30 gpu:3
gilbreth-e[000-015] 16 190000 E,e,V100,v100 gpu:2
gilbreth-f[000-004] 40 190000 F,f,V100,v100 gpu:2
gilbreth-g[000-011] 128 510000 G,g,A100,a100,A100-40GB,a100-40gb gpu:2
gilbreth-h[000-015] 32 512000 H,h,A10,a10 gpu:3
gilbreth-i[000-004] 32 512000 I,i,A100,a100,A100-80GB,a100-80gb gpu:2
gilbreth-j[000-001] 128 1020000 J,j,A100,a100,A100-80GB,a100-80gb gpu:4
To run your jobs in specific nodes, you can use -C, --constraint to specify the features. Below are a few examples:
#SBATCH --constraint 'E|F' ## request E or F nodes
#SBATCH --constraint A100 ## request A100 GPU
#SBATCH -C "v100|a100|a30" ## request v100, a100 or a30