Benchmarks TODO

[mratsim]

Important benchmarks

• Framework overhead via fibonacci
• Unbalanced Tree Search
• GEMM / Matrix Multiply vs OpenBLAS and MKL
• Binary size overhead when runtime is not compiled in
• Space overhead at runtime vs serial code
• Returning memory to the OS on long-running processes
• PARSEC benchmark suite: https://parsec.cs.princeton.edu/
• NAS Parallel Benchmarks from the NASA Advanced Supercomputing: https://www.nas.nasa.gov/publications/npb.html

Instrumentation, tutorials, examples

• topology: hyperthreading siblings, NUMA
• measuring performance, core usage, latencies, cache misses, view assembly:
  • perf
  • Intel VTune
  • Apple Instruments
• bloaty for binary size
• perf c2c for measuring cache contention / false sharing
• helgrind for locking

Requires changing the internals:

• coz for causal profiling and bottleneck detection
• relacy for race detection

Stretch goals

• Other common benchmarks (nqueens, nbodies, LU, heat, qsort, bouncing producer-consumer, ...)
• Porting michi (550 lines go bot with parallel Monte-Carlo Tree Search in Python) to Nim (https://github.com/pasky/michi) and benching against the C and Go implementations.