In order to test performance portability across a wide range of devices, we maintain a small cluster containing many different accelerator technologies.

Access is by request.

Contents

Name Class Architecture Queue command
NVIDIA K20m HPC GPU Kepler -lnodes=1:gpus=1:k20
NVIDIA K40m HPC GPU Kepler -lnodes=1:gpus=1:k40
NVIDIA GTX 580 Consumer GPU Fermi -lnodes=1:gpus=1:gtx580
NVIDIA GTX 680 Consumer GPU Kepler -lnodes=1:gpus=1:gtx680
NVIDIA GTX 780 Ti Consumer GPU Kepler -lnodes=1:gpus=1:gtx780ti
NVIDIA GTX 980 Ti Consumer GPU Maxwell -lnodes=1:gpus=1:gtx980ti
NVIDIA GTX 1080 Ti Consumer GPU Pascal -lnodes=1:gpus=1:gtx1080ti
NVIDIA GTX TITAN X Consumer GPU Pascal -lnodes=1:gpus=1:titanx
AMD S9150 HPC GPU Hawaii -lnodes=1:gpus=1:s9150
AMD S10000 HPC GPU Tahiti Unavailable
AMD HD7970 Consumer GPU Tahiti -lnodes=1:gpus=1:hd7970
AMD R9-295X2 Consumer GPU Hawaii -lnodes=1:gpus=1:r9-295x2
AMD R9-290X Consumer GPU Hawaii -lnodes=1:gpus=1:r9-290x
AMD Fury X Consumer GPU Fiji -lnodes=1:gpus=1:furyx
AMD RX 480 Consumer GPU Polaris -lnodes=1:gpus=1:rx480
Intel Xeon E5-2697 v2 Server CPU Ivy Bridge -lnodes=1:ppn=24:ivybridge
AMD A10-7850K Radeon R7 APU Kaveri -lnodes=1:kaveri
Intel Xeon Phi 7210 MIC KNL -lnodes=1:ppn=256:knl
NVIDIA Jetson TX1 ARMv8 Cortex-A57 -lnodes=1:ppn=4:jetson
SoftIron Overdrive 1000 ARMv8 Cortex-A57 -lnodes=1:ppn=4:overdrive

When running on the KNL, please see additional documentation.

Documentation

Connect to the headnode via ssh at hpc.cs.bris.ac.uk.

Software

Tools and compilers are made available via the module command.

List available software: module avail

Load a module: module load <name>

Unload a module: module unload <name>

MPI

We have a selection of MPI builds installed for a variety of circumstances. It is recommended to have only one MPI module loaded at once to prevent interference.

The Intel modules include the Intel MPI library.

OpenMPI is built in a number of flavours. Unless a compiler version is specified they use the default GCC v4.8.4 compilers.

Running jobs

We use the TORQUE (PBS) queueing system to manage job allocations of the cluster.

List the nodes and their current availability by running

znodes

Interactive job

You can launch an interactive job requesting the gpu <name> (all lower case) using

qsub -lnodes=1:gpus=1:<name> -I

You can specify any NVIDIA or AMD gpu using the following:

qsub -lnodes=1:gpus=1:nvidia -I
qsub -lnodes=1:gpus=1:amd -I

You can also request by architecture, for example:

qsub -lnodes=1:gpus=1:kepler