1. dalma-system
2. Dalma Manual
3. dalma-system

Dalma System Configuration

SYNOPSIS

You want to know the system configuration of Dalma, in order to make the most out of it.

NOTICE

Bigmem, GPU and Visualization nodes will be updated.

DESCRIPTION

The most important specification are, the number of cpus per normal compute node: 28, and memory per node: 128GB. Detailed configuration for normal compute nodes:

• 236 compute nodes
• Two Intel Xeon E5-2680 v4 Broadwell CPUs per node
• 2 x 14 cores, 2.4 GHz
• No Hyperthreading
• AVX 2.0 ISA extension
• 128 GB Memory Per Node
• 6,608 CPU cores
• Mellanox EDR InfiniBand
• CentOS 7 Linux distribution
• Slurm batch system
• Support for OpenMP programming model for intra-node parallelization
• OpenMPI (Message Passing Interface) Implementation
• BeeGFS Parallel Filesystem for regular files
• Lustre Parallel Filesystem for large files

We provide Fat, GPU nodes for extra computing requirements. Contact us to tailor your job script.

EXPLANATION

A long list. So what does it mean?

• Job script should be written in SLURM.
• For parallel jobs, utilize the whole node (28 cores) to maximize the performance. E.g., set OMP_NUM_THREADS for hybrid or OpenMP jobs, or set the number of MPI ranks equals the number of nodes times 28 for pure MPI jobs.
• Use maximum 100GB of memory per node. If you need more, use fat node. Why not 128GB? We need to reserve memory for system itself and etc.
• For pre-compiled binaries, try Centos or RHEL version.
• Before generating tons of data, evaluate your storage requirements and select the most suitable file system.

AUTHORS

NYUAD HPC Apps Team:
--------------------
    - Benoit Marchand
    - Guowei He
    - Jorge Naranjo

SEE ALSO

Please refer to the online documentation available here

1. NYUAD
2. February 2017
3. dalma-system