1 / 42

Implications of Classical Scheduling Results For Real-Time Systems

Implications of Classical Scheduling Results For Real-Time Systems. John A .Stankovic ,Marco Spuri ,Marco Di Natale ,and Giorgio Buttazzo Present : Bo- yuan Chen 陳柏元. outline. Introduction Static versus Dynamic Scheduling Metrics Uni-processor Systems Preemption and NonPreemption

aron
Download Presentation

Implications of Classical Scheduling Results For Real-Time Systems

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Implications of Classical Scheduling Results For Real-Time Systems John A .Stankovic ,Marco Spuri ,Marco Di Natale ,and Giorgio Buttazzo Present:Bo- yuan Chen 陳柏元

  2. outline • Introduction • Static versus Dynamic Scheduling • Metrics • Uni-processor Systems • Preemption and NonPreemption • Precedence Constraints • Shared Resources • Overload and Value

  3. Introduction • Important classical scheduling theory results for real-time computing are identified. • Implications of these results from the perspective of a real-time systems designer are discussed.

  4. Static versus Dynamic Scheduling • Static scheduling algorithm has complete knowledge regarding the task set and its constraints such as deadlines ,computation times, precedence constraints… • Dynamic scheduling algorithm has complete knowledge of the currently active set of tasks ,but new arrivals may occur in the future ,not known to the algorithm at the time it is scheduling the current set.

  5. Static versus Dynamic Scheduling • Off-line scheduling is often equated to static scheduling ,but this is wrong! • In building any real-time system , off-line analysis should always be done regardless of whether the final runtime algorithm is static or dynamic.

  6. Metrics • Minimizing the sum of completion times • Minimizing the weighted sum of completion times • Minimizing the number of processors required • Minimizing the maximum lateness

  7. Which of these metrics are interest to real-time systems designer? • The sum of completion times is generally not of interest because in many real-time systems where the focus is simply on deadlines. • Minimizing the maximum lateness metric is not always useful because minimizing the maximum lateness does not prevent one or many tasks from missing their deadlines.

  8. d1 d1 d2 d2 d3 d3 d4 d4 d5 d5 The first schedule minimizes the max lateness ,but all tasks miss their deadline Maximum lateness The second schedule has a greater max lateness ,but four tasks out of five complete before their deadlines Maximum lateness

  9. Uni-processor Systems • Problem definition: • α|β|γ , whereαindicates the machine environment , α=1 indicating a uni-processor machine ; • βindicates the job characteristics (preemptable , nonpreemptable ,independent ,precedence constrained , etc.) ; • γindicates the optimality criterion (maximum lateness)

  10. Preemption and NonPreemption • Job j having a processing time pj and a due date dj ;defined completion time Cj ,lateness of j is Lj = Cj – dj ; Lmax=max{Lj} • problem: 1|nopmtn|Lmax ;objective function to minimize is Lmax

  11. Preemption and NonPreemption • This is a static scheduling algorithm • If the maximum lateness is greater than zero , then the designer knows that he must increase the computing power of his system to meet the requirements of missing no deadlines • Note that all tasks are known and ready to execute at time zero , preemption would not improve the situation

  12. Preemption and NonPreemption • If our real-time system requires to consider release times rj ,the problem is • 1|nopmtn, rj |Lmax • Unfortunately ,the problem is NP-hard • But 1|pmtn, rj | Lmax has polynomial complexity

  13. Preemption and NonPreemption • Preemption usually gives greater benefit than nonpreemption in terms of scheduling complexity.

  14. Preemption and NonPreemption • For large n we obtain the utilization bound of 69% meaning that as long as the CPU utilization is less than 69% all tasks will make their deadlines.

  15. Precedence Constraints • Precedence constraints will be expressed with the notation i j , or with their associated digraph G(V,E) where V is the set of tasks and E the set of edges ,an edge connecting tasks i ,j if task i precedes task j.

  16. Precedence Constraints • Problem:1|prec,nopmtn| Lmax • Solve: Lawler EDF-like algorithm • Scheduling list is built starting from the bottom in reverse topological order • The task having the minimum value for the chosen metric and whose successors have been scheduled • Adding to the list on each step • Lawler’s algorithm is optimal

  17. Precedence Constraints • Problem:1|nopmtn,rj | Lmax and 1|prec,nopmtn,rj| Lmax proven to be NP-hard by Lenstra.

  18. Precedence Constraints • Polynomial algorithms have been found for precedence relations in the form: • Intrees: every task has no more than one predecessor • Outtrees: task have no more than one sucessor • series-parallel graph

  19. Precedence Constraints • series-parallel graph: • G({j},0) is a series-parallel graph • If G1(V1,A1) and G2(V2,A2)are series-parallel graphs than • G1G2= (V1∪V2 , A1∪A2∪(V1× V2))and • G1||G2 = (V1∪V2 , A1∪A2) are series-parallel graphs

  20. Precedence Constraints • A graph is a series-parallel graph only if its transitive closure does not contain the Z graph • Z graph is a graph that contains as a subgraph 4 nodes(i,j,k,l) with only the follow edges • ij , ik , lk

  21. Precedence Constraints • Program example that gives rise to Z graph

  22. Precedence Constraints • Problem:1|prec,pmtn,ri|Lmax • Baker’s procedure can solve the problem • Number of preemption is limited to n-1 where n is the number of jobs • Preemption overhead bounded • We rarely see this issue addressed in classical scheduling theory

  23. Precedence Constraints • Encode the precedence relations into the parameters used by the scheduling algorithm , for example , into deadlines and release times.

  24. Precedence Constraints

  25. Precedence Constraints • Blazewicz’s result comes from the fact that task deadlines depend on their deadlines and successors’ deadlines , while task start times depend on their own start time and predecessors’ start times • The theorem assumes no shared resources among tasks • This result allows us to transform a set of dependent tasks into a set of independent ones obtaining an equivalent problem under the EDF policy.

  26. Precedence Constraints • The optimality of the technique of the revised deadlines and arrival dates has been used in both on-line. • Unfortunately ,the general problem of scheduling a set of tasks with precedence constraints and arbitrary resource conflicts is NP-hard.

  27. Shared Resources • In Mok’s opinion ,The reason for the NP-hardness of the above scheduling problem lies in the possibility that there are mutually exclusive scheduling blocks which have different computation times

  28. Shared Resources • A confirmation of this point of view is that the problem of minimizing the maximum lateness of n independent unit-time jobs with arbitrary release times ,1|nopmtn,rj,pj=1|Lmax is easy.

  29. Shared Resources • Moreover , if we add precedence constraints and we want to minimize the maximum completion time : • 1|nopmtn ,prec , rj, pj=1|Cmax is still easy.

  30. Shared Resources • Priority ceiling protocol(PCP) ,an allocation policy for shared resources which works with a rate monotonic scheduler(fixed priority) • The key ideas behind the PCP is to prevent multiple priority inversion by means of • early blocking of tasks that could cause priority inversion • Minimize as much as possible the length of the same priority inversion allowing a temporary rise of the priority of the blocking task

  31. Shared Resources • PCP: • Define the ceiling of a critical section as the priority of the highest priority task that currently locks or could lock the critical section • Allow the locking of a critical section only if the priority of the requesting task is higher than the ceiling of all critical sections currently locked • In case of blocking , the task that holds the lock inherits the priority of the requesting task until it leaves the critical section

  32. Shared Resources • PCP properties: • A job can be blocked at most once before it enters its first critical section • The PCP prevents the occurrence of deadlocks

  33. Shared Resources • Comparison: PIP( Priority Inheritance protocol) and PCP • PCP is more efficient at run-time, in that a high priority task cannot be blocked more than once before it enters its first critical section • PIP does not require a thorough analysis of a task’s behavior (in terms of which semaphores in might request)

  34. Shared Resources • Stack Resource Policy (SRP) : • Similar to PCP, • PCP: Task is blocked when it wants to lock resource • SRP: Task is blocked when it attempts to preempt • Multiunit resources • Static and dynamic priority schemes • Sharing of runtime stacks

  35. Shared Resources • SRP relies on the following two conditions: • To prevent deadlocks ,a job should no be permitted to start until the resources currently available are sufficient to meet its maximum requirements. • To prevent multiple priority inversion ,a job should not be permitted to start until the resources currently available are sufficient to meet the maximum requirement of any single job that might preempt it.

  36. Shared Resources • SRP advantages: • Earlier blocking are to save unnecessary context switch • Simple and efficient implementation of the SRP by means of a stack

  37. Overload and value • EDF and LLF ,in overload conditions ,these algorithms perform very poorly. • Domino effect: since the first task that misses its deadline may cause all subsequent tasks to miss their deadlines. • A value is usually associated with each task , reflecting the importance of that task within the set. Tasks can be scheduled by the Smith’ rule.

  38. Overload and value • Similar to shortest processing time first (SRT). • Is equivalent to SRT when all tasks have equal weights.

  39. Overload and value • Baruah have shown that there exists an upper bound on the performance of any on-line(preemptive)algorithm working in overload conditions

  40. Overload and value • Definition: • Clairvoyant scheduler: one that knows the future • Each task assign a value: • Equal to the task’s execution time if task request is successfully scheduled to completion • Equal to zero if task that do not terminate within their deadline. • Competitive factor: cumulative value achieved by the algorithm to the cumulative value achieved by the clairvoyant schedule.

  41. Overload and value • These bounds are true for any load. • Above bound is achieved under very restrictive assumption.In most real world applications tasks characteristics are much less restrictive. • 1/4 th bound has only a theoretical validity and more work is needed to derive other bounds based on more knowledge of the task set.

  42. Conclusion • Many basic algorithms are based on earliest deadline scheduling or rate monotonic scheduling. • Designers of real-time systems have a wealth of information concerning uni-processor scheduling.

More Related