1. Limited-Preemption Scheduling of Sporadic Tasks Systems
RETIS Lab
Real-Time Systems Laboratory
Research Area: Real-Time Scheduling and Resource Management
Limited-Preemption Scheduling of Sporadic Tasks Systems
Marko Bertogna

2. Sporadic task system with arbitrary deadlines
Introduction
Sporadic task system with arbitrary deadlines
t = t1, t2,…, tn with ti = (ei ,di ,pi)
Preemptive EDF is an optimal scheduler
Exact feasibility test:
with
for each , until a pseudo-polynomially
far point

3. To preempt or not? PREEMPTIVE Optimal schedulability performances
Need to use protocols for the access to shared resources
NON-PREEMPTIVE
Higher feasibility overhead
Lower run-time overhead
Simplified access to shared resources
Ideal situation: optimal scheduling algorithm
with low run-time overhead
Preemptive EDF is not written to minimize preemptions
EDF is an optimal scheduler for uniprocessor systems, but not the only one
Non-preemptive EDF:
Not optimal under non-preemptive model
Optimal under non-preemptive model without inserted idle times
Allow preemption only when necessary
for maintaining feasibility

4. Limited-preemption EDF
Non-preemption function Q(t)
Jobs priorities according to EDF
Two modes: regular and non-preemptive
Initially, a job JL executes in regular mode
When a higher priority job JH arrives, JL goes in non-preemptive mode JH
Regular t
min[cL,Q(DL - t)] DL JL
Regular
Non-Preemptive

5. Non-preemption function Q(t)
Compute Q(t) such that
Feasibility is maintained
Non-preemptive sections as large as possible
Properties of Q(t)
Monotonic non-increasing
Changes value only at time-instants corresponding to task deadlines in a synchronous periodic release sequence

6. For every deadline D2, D3, …, Dm ≡ dmax :
Computing Q(t)
For every deadline D2, D3, …, Dm ≡ dmax :
Computed only at tasks deadlines, until a pseudo-polynomial point.
Same operations as in the EDF feasibility
check:

7. Pseudo-polynomial complexity
Comes for free when feasibility has to be checked as well
When storing the Q(t) table, possible to discard some value, finding suboptimal results
Very small memory requirements (from simulations)
No more than 9 points of discontinuity
Average number of 3 discontinuities

8. Simulations uniform Ui n = 5 pi in [10,1000] t in [0,106]
For each (n,U) 1000 task sets.
3 to 5 times lower

9. Simulations uniform Ui n = 10 pi in [10,1000] t in [0,106]
For each (n,U) 1000 task sets.
5 to 10 times lower
Always les than preemptions every 1M time-units for every n and U.
Independent from n!!! While preemptive EDF proportional to n!!!

10. Considerations and conclusions
Optimal scheduling algorithm based on EDF
Reduced number of context changes
Small computational complexity and memory requirements
Advantages w.r.t. preemptive EDF
Lower run-time overhead
Easy way to deal with shared resources
Enhanced predictability

11. Marko Bertogna PhD student marko@sssup.it
RETIS Lab
Real-Time Systems Laboratory
Thank you
Marko Bertogna PhD student