1. Missed Deadline Notification in Best-Effort Schedulers Balaji Raman

2. Outline Motivation Methodology Implementation Details Results Conclusion

3. Motivation – What are Best Effort Schedulers? Best effort scheduling Does not provide facilities to specify or meet performance guarantees. Advantages –Simple –Applications do not require system interfaces Suitability –Soft real time systems Disadvantages –Not suitable for hard real time systems –Under Utilization

4. Methodology-MDN Indicate deadline misses to scheduler. Problems in all existing general purpose system schedulers- Linux, BEST.. Advantages of MDN Minimal interface

5. Implementation Details Implemented in three schedulers Linux Scheduler BEST Scheduler Be-Rate Scheduler We will look MDN implementation in Linux Scheduler in detail.

6. Linux Time-Share Scheduler Scheduling based on Quantum Counter Variable A process can block its task if it is periodic. Scheduler Q=1 C=5 Tid=A Q=2 C=1 Tid=B Q=3 C=1 Tid=C Q=4 C=1 Tid=D Scheduler Q=1 C=1 Tid=B Q=2 C=1 Tid=C Q=3 C=1 Tid=DQ=4C=4Tid=A

7. Linux Time-Share Scheduler When all counters in queue are zero Counter=Counter/2+nice Problem:Problem: Deadlines could be missed for tasks with counter=0. Scheduler Q=1 C=0 Tid=B Q=2 C=0 Tid=C Q=3 C=0 Tid=DQ=4C=1Tid=A

8. MDN + Linux Counter=counter/2+(nice+mdn_ nice) –Initially mdn_nice=0. –If MDN received mdn_nice++ –Priority boosted for not missing future deadlines. Anti-Cheat Policy

9. Best Scheduler Detect Periodic tasks –Improves performance of soft real- time tasks Tasks entering Run Queue Run Q Period of the task ~ Y-X Scheduler (Task D Entering Times: Rec_Time=X Prev_Time=Y) Q=1 Tid=A Q=2 Tid=B Tid=DCurr_Time=X

10. Best Scheduler Scheduling Policy Earliest Deadline First. Problems BEST Detects Period Deadline

11. BEST + MDN BEST Detects Period Deadline MDN: Deadline Missed BEST Detects Period MDN Detects Misses in deadline. Anti-cheat Policy Based on block ratio

12. Be-Rate Scheduler Rate based algorithm. MDN Enhancement. Anti-Cheat Policy.

13. Experimental Evaluation Measure performance by observing the number of deadlines missed with & without MDN. Benchmark: 2 synthetic workload application CPU bound, Soft real-time process. Experiments done on Intel Pentium P3 and P4 processors.

14. Results (Linux) Time (seconds) Progress (CPU seconds) xxxxxxx xx xx xxx xx xxxxxxxx xxxx xxxx xxxxxx CPU-bound SRT-50% xx: Deadline Missed

15. Results (Linux + MDN) Time (seconds) Progress (CPU seconds) xx x CPU-bound SRT-50% xx: Deadline Missed

16. Results (Linux) Time (seconds) Progress (CPU seconds) xxxxxxx xx xx xxx xx xxxxxxxx xxxx xxxx xxxxxx CPU-bound SRT(25)-33% xx: Deadline Missed SRT(33)-33% xx: Deadline Missed xxxxxxx xx xx xxx xx xxxxxxxx xxxx xxxx xxxxxx

17. Results (Anti Cheating Policy) Time (seconds) Progress (CPU seconds) CPU-bound Cheating Process

18. Results (Linux + MDN) Time (seconds) Progress (CPU seconds) CPU-bound SRT(25)-33% xx: Deadline Missed SRT(33)-33% xx: Deadline Missed xx x

19. Conclusion Best-effort schedulers an attractive advantage. MDN makes Best-effort schedulers more efficient. Future work Tuned to workloads. Rate controlled MDN.

20. Back Up Slides BeRate –Rate based algorithm –Two Parameters Task should complete e units of computation over p interval of time. –Schedules so that they can continue processing in the same rate as in the past. –Uses EDF scheduler.

21. Back Up slides Berate Scheduler –Given a fair share of processing Berate estimates period of the task so that task would meet its deadline using EDF scheduler –MDN is added. If estmated deadline is earlier then the task will be serviced quicker.

22. Backup Slides Results of BEST, BeRate –BeRate performance improvement slighter –Best tasks misses many deadlines because of overload, Best cannot detect period but using MDN it is able to alleviate its effect.