fupermod: Functional Performance Models of heterogeneous processors

Partitioner

Partitions 1D workloads. Takes conf_file and total problem size as input. Reads in performance model data files. Using 1 of 7 algorithms partitions workload to balance load. Outputs a text file partition.dist with hostname, device rank and distribution for later input into routine (or fupermod_1dto2d_mxm_dist).

Algorithms:

  1. Partial functional performance model
  2. Full functional performance model
  3. Constant performance model 1 (small kernel)
  4. Constant performance model 2 (homogeneous kernel)
  5. Homogeneous distribution
  6. Manual distribution
  7. Geometric Partitioning