Sprint 4

16/09 - 23/09

Agenda

1. Framework updates
2. Practical evaluation
3. Theoretical evaluation

Framwork updates

Performance
improvements

• Cache multiple collections in graph implementation
• Use sorted tree structures to reduce search time: log(n)
• Improved performance by factor ~20 for edge access and searching
• Still much potential for nodes and other search functions

Multiprocessor
improvements

• Improved task graph set scheduling
• Split load into multiple threads
• One thread for each task graph set and each scheduler
• Still potential for asynchronous writing and reading of files

Practical evaluation

Approach

• Execute given schedule
• One Java thread (worker) for each cpu
• Check dependencies instead of startTime
• Synchronized results list across all workers
• Perform fake work with time scale of 1 : 20
  • Time based worker
  • Load based worker
• Returns schedule of simulated execution

Example

Summary

• Practical execution takes longer than theoretical values
• There are some more gaps because of small timing differences
• Timbased approach is closer to theoretical one than the loadbased one
• Loadbased approach is strongly influenced by system setup and other load on the system

Theoretical evaluation

Structure

• Graph overview
• Group overview
• Scheduler overview
• Several diagramtypes
• Evaluation for 10, 50, 100, 300