Skip to content

Loss of performance when increasing the number of workers

To benchmark the code and find the optimal configuration, I tried to run one tile on one node (HAL). The configuration was 9 SAFE files (18 images). I reserved one full node on HAL (40 CPU and 100 Gb)

  • With nbthread=2 and nbworker=5:

AnalyzeBorder + Calibrate + Curborder = 68 seconds (average) Orthorectification = between 80 secs and 170 secs

  • With nbthread=2 and nbworker=20:

AnalyzeBorder + Calibrate + Curborder = 68 seconds (average) (no change) Orthorectification = between 485 secs and 700 secs

  • With nbthread=4 and nbworker=5:

AnalyzeBorder + Calibrate + Curborder = 35 seconds (average) (no change) Orthorectification = between 100 secs and 130 secs

So, something is going wrong on Orthorectification when using more workers. May be because a concurrent access to DEM data and/or Geoid file