blomma-graphs-logo

Sprint 5

23/09 - 26/09

Agenda

  1. Optimization
    1. LAST Scheduler
    2. DLS Scheduler
    3. Custom Stream Scheduler
  2. Overall results

Optimization

LAST Scheduler

  • Currently just appending -> gaps
  • Fill up gaps when scheduling new task
  • Consider dependencies and gap size


Problems

  • Small improvement for costs of long algorithm duration
  • Bypassing gaps if dependencies are not finished at beginning of gap
  • Error-prone and many special cases

DLS Scheduler

Custom Stream Scheduler

  • The next task to schedule just depends on the gap, deadline and EST
  • considering the communication time when choosing the next task
  • > could get more likely a higher throughput
  • > in the best case save the difference between the communication times

Summary
of results

  • gap idle time, waiting for dependencies
  • idle time during communication time
  • unnecessary communication time (better to schedule on same cpu)
  • save time with zeroing communication time by scheduling on the same cpu
  • distribute load to all of the cpus
  • create as less gaps as possible
  • keep within the deadlines

getting as close as possible to the CP while keeping within the deadlines