Load balancing key points

  • There is 1 overhead to a PUT
    • Performing the PUT
  • There are 2 overheads to a GET
    • Waiting for the data
    • Performing the GET
  • Every GET risks a model having to wait
  • Every timestep there is variation due to
    • differing numerics and
    • computer jitter
  • You want each model to run at roughly the same speed
    • However, you don't want your expensive models to wait