News from GPU parameter tuning

Changes after the mismatch fix

• Tuned PbPb 50 KHz after fixing the HIP/CUDA mismatch
• Lowered grid sizes to maximum of 600 (kept max 900 only for some small kernels)
• Lowered number of trials
• New constraint in the optimization function:
  if min_block_per_cu * max_threads / 256 > 10:
   return float("inf")
  Penalize configurations that would make the second parameter > 10
• In this way, avoid kernels with no register spillage (since the compiler ignores min_active_warps > 10) but with high grid size
• Forces the algorithm to focus on small but good configurations

• Best parameters perf [ms]	Before mismatch fix	After mismatch fix + constraints
  Clusterizer step	732.18	757.02
  CompressionKernels_step1unattached	523.69	524.31
  GMMergerCollect	93.86	94.40
  GMMergerSectorRefit	284.58	267.72
  GMMergerTrackFit	761.78	704.26
  FollowLoopers + CompressionKernels_step0attached	508.43	480.44
  TrackletConstructor step	1091.24	> 4000

For the tracklet constructor step, since it is highly dimensional, it was more probable to fall into the condition min_block_per_cu * max_threads / 256 > 10, leading to many inf values and few acceptable values. Removing the constraint lead to an optimal solution of 1128.05.

Reduced grid sizes in best parameters. Some examples:

NVIDIA GPUs tuning

Got access to the NGT cluster. Started to adapt the tuner for Nvidia GPUs.

ToDos:

• Investigate why big grid sizes with low occupancy are preferred rather than the small grid sizes with same occupancy
• Decide if how to integrate new default parameters in O2

ALICE contribution to HS23

• Created repository for the Dockerfile: https://github.com/cima22/O2GPU_HS23
• Waiting reply by Robin for testing