News from GPU parameters tuning

Robustness test

• Simulated 3 PbPb 50kHz datasets
• Set memSize 15 GBs
• Tuned params with each dataset --> 3 different best parameters dumps

Clusterizer step tuning

• Many static asserts
• Many thread counts need to be the same among kernels
• --> Constrained optimization
• Optuna should have soft constrained optimization
  • Can't force constrained, but objective function penalized if not satisfying that constraint
  • Already have that by default: if benchmark crashes / does not compile, backend returns "Inf" value
• Need to investigate how to deal with this