NODE BURN

Because sometimes burning the GPU is not enough.

A simple tool for running compute or memory intensive workloads on both CPU and GPU, in order to understand

• the performance of the individual components
• the impact of the workloads on one another
• the maximum power consumption of a node

Features

Node burn can run GEMM or STREAM workloads on the GPU only, CPU only, or both simultaneously.

# run GEMM with matrix dimension 5000*5000 on the GPU,
# and STREAM triad with length 500000 on the CPU, for 30 seconds.
./burn -ggemm,5000 -cstream,500000 -d30

# run GEMM on the GPU, nothing on the CPU, for 3 minutes.
./burn -cgemm,5000 -d180

# run GEMM on the CPU, nothing on the GPU, for 20 seconds.
./burn -ggemm,5000 -d20

Sometimes we want to run multiple instances of node burn in a parallel job, e.g. 4 instances on a node with 4 GPUs, or one instance on every GPU in a cabinet, to see if anything catches on fire understand system behavior under load. Use the --batch option to produce less verbose output that can be easily parsed by a post-processing script.

# on a system with 4 GPUs per node, use all 16 GPUs on 4 nodes to
# run GEMM with matrix dimension 10000*10000 on the GPU for 30 seconds
srun -n16 -N4 ./burn --batch -ggemm,10000 -d30
nid001272:gpu    584 iterations, 38930.92 GFlops,     30.0 seconds,    2.400 Gbytes
nid001272:gpu    579 iterations, 38555.99 GFlops,     30.0 seconds,    2.400 Gbytes
nid001272:gpu    561 iterations, 37348.14 GFlops,     30.0 seconds,    2.400 Gbytes
nid001272:gpu    600 iterations, 39939.47 GFlops,     30.0 seconds,    2.400 Gbytes
nid001278:gpu    585 iterations, 38994.35 GFlops,     30.0 seconds,    2.400 Gbytes
nid001278:gpu    584 iterations, 38914.98 GFlops,     30.0 seconds,    2.400 Gbytes
nid001278:gpu    589 iterations, 39200.59 GFlops,     30.1 seconds,    2.400 Gbytes
nid001278:gpu    589 iterations, 39204.37 GFlops,     30.0 seconds,    2.400 Gbytes
nid001274:gpu    557 iterations, 37091.74 GFlops,     30.0 seconds,    2.400 Gbytes
nid001274:gpu    560 iterations, 37289.96 GFlops,     30.0 seconds,    2.400 Gbytes
nid001274:gpu    542 iterations, 36090.85 GFlops,     30.0 seconds,    2.400 Gbytes
nid001274:gpu    503 iterations, 33473.36 GFlops,     30.1 seconds,    2.400 Gbytes
nid001276:gpu    584 iterations, 38929.67 GFlops,     30.0 seconds,    2.400 Gbytes
nid001276:gpu    589 iterations, 39253.24 GFlops,     30.0 seconds,    2.400 Gbytes
nid001276:gpu    588 iterations, 39170.08 GFlops,     30.0 seconds,    2.400 Gbytes
nid001276:gpu    589 iterations, 39224.21 GFlops,     30.0 seconds,    2.400 Gbytes

Power Measurement on HPE-Cray systems

If running on a HPE Cray-EX system with pm_counters, nodeburn can be configured to generate a report of power consumption on each node. Enable it at build time with the NB_PMCOUNTERS CMake option (see below).

node-burn will generate power reports from all of the energy counters that it can detect on each node - the values reported will vary according to the node architecture.

Requirements

C++20 for the C++ code, C++17 for the CUDA code.

It has only been tested with GCC 11+ and CUDA 11.8+.

• not tested with Clang, Intel, NVC, Cray compilers. It should work if the compiler is recent.

Compiling

Node burn uses CMake to configure the build. There is currently one option, NB_GPU which can be used to to disable CUDA targets.

# by default node burn will attempt to build for CUDA devices.
CC=gcc CXX=g++ cmake $src_path

# explicitly disable building for CUDA
CC=gcc CXX=g++ cmake $src_path -DNB_GPU=off

On HPE Cray-EX systems, power readings from pm_counters can be generated using the NB_PMCOUNTERS option.

# enable pm counters for average power consumption
CC=gcc CXX=g++ cmake $src_path -DNB_PMCOUNTERS=on