1. © COPYRIGHT IKERLAN 2014 Adding Precedence Relations to the Response-Time Analysis of EDF Distributed Real-Time Systems Unai Díaz de Cerio (IK4-Ikerlan) Michael González Harbour (Univ. Cantabria) J. Carlos Palencia (Univ. Cantabria) Juan P. Uribe (IK4-Ikerlan) 22nd International Conference on Real-Time Networks and Systems Versailles, France, October 8-10, 2014

2. © COPYRIGHT IKERLAN 2014 Motivation  Context ­ Distributed real-time systems ­ Response-time schedulability analysis ­ EDF scheduling: GC-EDF (Global-Clock EDF)  Global clock or clock synchronization available LC-EDF (Local-Clock EDF)  One clock per processor and no clock synchronization available  Problem ­ No exact solution for schedulability analysis in distributed systems ­ More techniques to reduce the pessimism in the analysis for FP (fixed- priorities) than for EDF 2

3. © COPYRIGHT IKERLAN 2014 Motivation Analysis techniqueFPGC-EDFLC-EDF 3 Tindell and Clark, 94Spuri, 96Rivas et al., 10Holistic WCDO Palencia and González Harbour, 98 Palencia and González Harbour, 03 Palencia and González Harbour, 99 WCDOPS XX X WCDO : Worst Case Dynamic Offsets WCDOPS: Worst Case Dynamic Offset with Precedence Schemes

4. © COPYRIGHT IKERLAN 2014 Objectives  Reduce the pessimism of the response time analysis for EDF scheduled systems ­ Extending dynamic offsets technique (WCDO) to LC-EDF scheduled systems ­ Extending precedence relations technique (WCDOPS) to LC-EDF and GC-EDF scheduled systems 4

5. © COPYRIGHT IKERLAN 2014 System model  Distributed system of n task/messages statically allocated in m processors/networks  Tasks/messages ( τ ij ) linked making up end-to-end flows ( Γ i )  Periodic end-to-end flows with minimum time between instances: T i  Task offset:  Task release jitter:  Task release time interval: ­ t = external event arrival time 5

6. © COPYRIGHT IKERLAN 2014 System model  Task worst case execution time (WCET): C ij  Relative global deadline (related to external event arrival): D ij  Relative local deadline (related to task release in its processor): d ij  Global response time (related to external event arrival): R ij  Local response time (related to task release in its processor): r ij 6 d i3 D i3

7. © COPYRIGHT IKERLAN 2014 WCDO-EDF (LC-EDF) Analysis techniqueFPGC-EDFLC-EDF 7 Tindell and Clark, 94Spuri, 96Rivas et al., 10Holistic WCDO Palencia and González Harbour, 98 Palencia and González Harbour, 03 Palencia and González Harbour, 99 WCDOPS XX X WCDO : Worst Case Dynamic Offsets WCDOPS: Worst Case Dynamic Offset with Precedence Schemes

8. © COPYRIGHT IKERLAN 2014 WCDO-EDF (LC-EDF)  Find worst case busy period ( length t and deadline D, created with the task τ ik ) and calculate the contribution of the tasks to this busy period.  Jobs categorized in 3 sets: ­ Set 0: Activations that occur before the busy period and that, even applying its maximum jitter, cannot be delayed until the busy period. ­ Set 1: Activations that occur before the busy period and can be delayed with an amount of jitter such that they coincide with the beginning of the busy period. ­ Set 2: Activations that occur inside the busy period. 8

9. © COPYRIGHT IKERLAN 2014 Jobs with deadline previous to D Number of jobs in Set2 Only jobs with deadline previous to D can contribute Number of jobs in Set1 WCDO-EDF (LC-EDF)  Maximum contribution when jobs in Set 1 suffer an amount of jitter that they are release at the beginning of the busy period and job in Set 2 suffer an amount of jitter equal to zero: 9

10. © COPYRIGHT IKERLAN 2014 WCDOPS-EDF (LC-EDF) Analysis techniqueFPGC-EDFLC-EDF 10 Tindell and Clark, 94Spuri, 96Rivas et al., 10Holistic WCDO Palencia and González Harbour, 98 Palencia and González Harbour, 03 Palencia and González Harbour, 99 WCDOPS XX WCDO : Worst Case Dynamic Offsets WCDOPS: Worst Case Dynamic Offset with Precedence Schemes

11. © COPYRIGHT IKERLAN 2014 Precedence relation and activation conflicts 11

12. © COPYRIGHT IKERLAN 2014 Precedence relation and activation conflicts 12

13. © COPYRIGHT IKERLAN 2014 Precedence relation and activation conflicts 13  Activation conflicts: Two tasks are in conflict when their executions are incompatible in the same busy period for the purpose of analyzing task τ ab.  H-section: Two tasks τ ij and τ ik are in the same H-section, for the analysis of τ ab, if both execute in the same processor as τ ab with priority equal to or higher than τ ab and there are no intermediate tasks between them in the same processor and with priority lower than τ ab. H ij ( τ ab ) identifies the H-section to which τ ij belongs.

14. © COPYRIGHT IKERLAN 2014 Applicability for EDF  Priority depends on absolute deadlines  Absolute deadline change from one job to the next  Contrary to fixed priorities  Activation conflicts between jobs, instead of between tasks  H-sections composed by jobs, instead of tasks 14

15. © COPYRIGHT IKERLAN 2014 Applicability for GC-EDF 15 Not useful  Task absolute deadlines are related to the external event arrival  Task deadlines ordered D i1 < D i2  No activation conflicts

16. © COPYRIGHT IKERLAN 2014 Applicability for LC-EDF 16  Task absolute deadlines are related to the task release  Task deadlines are not ordered  There are activation conflicts Applicable

17. © COPYRIGHT IKERLAN 2014 WCDOPS-EDF (LC-EDF) 17  Contribution of each task: ­ First job that can be delayed to the busy period: ­ Last job in the busy period that have its deadline previous to the deadline of the task under analysis:

18. © COPYRIGHT IKERLAN 2014 WCDOPS-EDF (LC-EDF) 18  Construct an activation conflict table for the contribution of the end-to- end flow Γ i t1t2t3t4t5 p=-3C1C1 00C4C4 0 p=-2C1C1 C2C2 0C4C4 0 p=-1C1C1 C2C2 0C4C4 C5C5 p=0C1C1 C2C2 C3C3 C4C4 C5C5 p=1C1C1 0C3C3 C4C4 C5C5 p=2C1C1 0C3C3 00 W i (p=-3) = max(C 1, C 4 ) W i (p=-2) = max(C 1 +C 2, C 4 ) W i (p=-1) = max(C 1 +C 2, C 4 +C 5 ) W i (p=0) = C 1 +C 2 +C 3 +C 4 +C 5 W i (p=1) = max(C 1,C 3 +C 4 +C 5 ) W i (p=-3) = max(C 1, C 3 ) W i = W i (p=-3) + W i (p=-2) + W i (p=-1) + W i (p=0) + W i (p=1) + W i (p=2) = max(C 1, C 4 ) + max(C 1 +C 2, C 4 ) + max(C 1 +C 2, C 4 +C 5 ) + C 1 +C 2 +C 3 +C 4 +C 5 + max(C 1,C 3 +C 4 +C 5 ) + max(C 1, C 3 )

19. © COPYRIGHT IKERLAN 2014 Simulation results 19  Extensive simulations with random task sets generated with GEN4MAST [Rivas et al., 13] and analysed with MAST  Compared holistic technique [Rivas et al., 10] with the new algorithms WCDO-EDF and WCDOPS-EDF for LC-EDF distributed systems

20. © COPYRIGHT IKERLAN 2014 Simulation results 20  Processor maximum utilization:  Around 9% better utilization for WCDO-EDF  Around 16% better utilization for WCDOPS-EDF

21. © COPYRIGHT IKERLAN 2014 Conclusions and future work Analysis techniqueFPGC-EDFLC-EDF 21 Tindell and Clark, 94Spuri, 96Rivas et al., 10Holistic WCDO Palencia and González Harbour, 98 Palencia and González Harbour, 03 Palencia and González Harbour, 99 WCDOPS Not useful WCDO : Worst Case Dynamic Offsets WCDOPS: Worst Case Dynamic Offset with Precedence Schemes

22. © COPYRIGHT IKERLAN 2014 Conclusions and future work  Reduced significantly the pessimism of the response time analysis for systems scheduled by LC-EDF.  Improvements come “for free”. Only affect the analysis, not the scheduling policy, nor the system’s architecture.  Some works have slightly improved the WCDO and WCDOPS analysis for fixed-priorities (Mäki-Turja and Nolin, 08; Redell, 04). We have not taken into account for our analysis.  Future work: ­ Analyse previous improvements to try to reduce even more the pessimism. ­ Analyse the possibility to develop an exact test that will allow us to quantify the pessimism of these analysis. 22

23. © COPYRIGHT IKERLAN 2014 Bibliography [Palencia and González Harbour, 98] J. C. Palencia and M. González Harbour. Schedulability analysis for tasks with static and dynamic offsets. In Proc. 19th IEEE Real-time Systems Symp., pages 26–37, 1998. [Palencia and González Harbour, 99] J. C. Palencia and M. González Harbour. Exploiting precedence relations in the schedulability analysis of distributed real-time systems. In Proc. 20th IEEE Real-Time Systems Symp., pages 328–339, 1999. [Palencia and González Harbour, 05] J. C. Palencia and M. González Harbour. Response time analysis of EDF distributed real-time systems. J. Embedded Comput., 1(2):225–237, 2005. [Rivas et al., 10] J. M. Rivas, J. J. Gutiérrez García, J. C. Palencia, and M. González Harbour. Optimized deadline assignment and schedulability analysis for distributed real-time systems with local EDF scheduling. In Proc. 8th Int. Conf. on Embedded Systems and Applications, 2010. [Spuri, 96] M. Spuri. Holistic analysis for deadline scheduled real-time distributed systems. Technical report, INRIA, 1996. [Tindell and Clark, 94] K. Tindell and J. Clark. Holistic schedulability analysis for distributed hard real-time systems. Microprocessing and microprogramming, 40(2), 1994. [Rivas et al., 13] J. M. Rivas, J. J. Gutiérrez, and M. González Harbour. Fixed priorities or edf for distributed real- time systems? SIGBED Rev., 10(2):21–21, 2013. [Mäki-Turja and Nolin, 08] J. Mäki-Turja and M. Nolin. Efficient implementation of tight response-times for tasks with offsets. Real-Time Systems, 40(1):77–116, 2008. [Redell, 04] O. Redell. Analysis of tree-shaped transactions in distributed real time systems. In Proc. 16 th Euromicro Conf. on Real-Time Systems, 2004., pages 239–248. 23

24. © COPYRIGHT IKERLAN 2014 P.º J.M. Arizmendiarrieta, 2 20500 Arrasate-Mondragón (Gipuzkoa) Tel.: 943 71 24 00 Fax: 943 79 69 44 www.ikerlan.es 24