Discussion

• Disk parallelism vs. locality

  • high communication volume

• Computational load imbalance

  • non-uniform data distribution

• Possible solutions

  • use two-phase data composition
  • increase data block size