Introduction

• Aim of granularity control: change parallel execution to sequential execution or vice-versa based on conditions related to grain size and overheads.

• Benefits greater for systems with greater parallel execution overheads.
  • Small shared memory multiprocessors: beneficial.
  • Distributed memory multiprocessors, workstation ``farms'', etc: essential.

• In general cannot be done completely at compile-time: work done by a call to a recursive procedure often depends on the size of its input.

• Basic approach (DLH90):
  • Do as much work at compile-time as possible and relegate some tests and final decisions to run-time.

  • Generate at compile-time cost functions which estimate grain size as a function of input data size: evaluated at run-time when size is known.

  • Transformed programs automatically control granularity.