Controlling GPU auto-boost

Controlling GPU auto-boost is obsolete as of LSF 10.1 Fix Pack 4 because LSF now synchronizes with GPU auto-boost to resolve any issues that previously required disabling the auto-boost.

NVIDIA K80 hardware provides the auto-boost feature by default. Auto-boost is designed for applications that use a single GPU. For multi-GPU jobs, auto-boost results in different GPUs running at different speeds. LSF can turn off the auto-boost for jobs with exclusive thread/process multi-GPU requirements.

Set the LSB_GPU_AUTOBOOST parameter in the lsf.conf file to control this feature.

If set to N, LSF disables auto-boost for jobs that require multiple exclusive mode GPUs whether these GPUs are on one host or multiple hosts.

When a job requires shared mode GPUs and exclusive mode GPUs at the same time, LSF does not turn off auto-boost for this type of job. When this job is done or exited, LSF restores the original auto-boost status.

Note: LSF ignores an auto-boost configuration change while the sbatchd daemon is running. Because of NVML library API limitations, the sbatchd daemon cannot call the NVML library periodically to get the newest auto-boost configuration. If you need to change the GPU auto-boost configuration on one host, clean all the jobs on the host, modify the GPU auto-boost configuration, then restart sbatchd.

NVIDIA K80 hardware supports auto-boost. The NVML library must be Version 6.340 or newer.