LSB_GPU_POWEROFF_DURATION

Syntax

LSB_GPU_POWEROFF_DURATION=seconds

Description

A GPU consumes significant power even when it idles. LSF can decrease the GPU power consumed if GPU is not in use within a specified time. LSB_GPU_POWEROFF_DURATION enables this feature to specify the minimum time before LSF can power off a GPU if the GPU is idle.

LSF tries to allocate the GPU that is not running as "MIN power limit" mode. If the number of GPUs that are in "MAX power limit" mode are not enough, LSF allocates the GPUs that are in "MIN power limit" mode and switches those GPUs to run in "MAX power limit" mode.

LSF allocates the GPU flow as follows:
  • GPUs are in the same PCI.
  • Check if the MAX power limit mode GPUs meets job requirements. If they do, LSF does not allocate the MIN power limit mode GPUs first. If they do not meet the requirements, LSF allocates all the GPUs to the job, including both MAX power limit and MIN power limit mode GPUs.

If sbatchd is restarted, the GPU idle time is recalculated.

Specify a non-negative value as the minimum number of seconds that a GPU is idle before LSF powers it off.

If set to 0, LSF powers off GPUs immediately after the job finishes.

Nvidia K80 hardware supports auto boost and switch power limits. The NVML library must be Version 6.340 or newer.

Default

Not defined. LSF never powers off a GPU even when it is idle.