Code and Data Partitioning for Fine Grain Parallelism

1
Code and Data Partitioning for Fine Grain
Parallelism

• Michael Chu and Scott Mahlke
• University of Michigan
• November 6, 2015

2
Compiling in the Multicore Era

• Coarse-grain vs fine-grain parallelism

3
Objectives of this Work

• Goal detect and exploit available fine-grain
  parallelism
• Compiler support is key for good performance
• Divide computation operations and data across
  clusters
• Maximize direct access to values you need

4
Code and Data Partitioning

• Phased ordered approach
• Partition data first
• Computation second
• Profile-driven data partitioning
• Reduce stall cycles
• Coherence traffic
• Conflicts/misses
• Map computation partitioning to data location