Skip to main content
Link
Menu
Expand
(external link)
Document
Search
Copy
Copied
Search cs492-stevens
Lectures
Lecture 17 Template
17 - Scheduling (4)
Outline
Real-Time Scheduling
Schedulability
Real-Time Algorithms
Rate Monotonic Scheduling
Earliest Deadline First
Announcements
Reading:
Real-Time Scheduling
Schedulability
Rate Monotonic Scheduling
Earliest Deadline First
Evaluating Scheduling Algorithms
Typical Situation:
Scheduling Goals:
Categories:
Assumption:
For task k:
Tasks can be periodic or sporadic
A set of periodic tasks is schedulable <==>
A set of periodic tasks is schedulable
<==> a feasible schedule exists
(a schedule in which no task misses its deadline)
periodic tasks
is utilization of a single task k
Where:
Is this system schedulable?
If we can add a 4th task, what would its maximum
allowable processing time be?
Single-core CPU, periodic tasks.
is Utilization in single-core
Utilization is written with different
variables in the book:
Release Time: earliest time task can begin
Processing Time: expected execution time
Deadline: time processing must finish by
Period: time to next deadline = release time
PA 3 instructions still pending a few things.
currently due in 2 weeks (same day as midterm)
PA 2 due today
1)
2)
3)
4)
5)
MOS 2.5
-
-
-
-
-
-
-
- Patient vital monitoring
- Aircraft controls (fly-by-wire)
- Industrial automation controls
-
Physical device(s) generate events
Deadlines
Hard Real-Time: missed deadlines unacceptable
Software is written specially for easy analysis
Tasks run regularly, for limited time
no malloc, no recursion - provable upper limits on resource usage
sporadic = aperiodic, deadline not known until release.
(not in this class)
Task behavior is predictable and known in advance
Soft Real-Time: missed deadlines are tolerable
Predictability
Computer must react within fixed time
Reacting too late can be as bad as not reacting!
e.g.
100
50
30
100
200
500
Variants:
Fixed Priority - priorities are precomputed
Dynamic Priority - priorities are recomputed
-
-
-
-
-
-
-
-
-
-
Cyclic Executive (nonpreemptive superloop)
Earliest Deadline First (EDF)
Least Laxity First (LLF)
Simple algorithm, easy to implement
Lower schedulable utilization:
Short period always preempts long period tasks
Does not account for next deadline!
long period tasks delayed/preempted can cause missed deadlines
long period tasks often delayed/preempted can cause missed deadlines
Utility-based scheduling
Rate Monotonic (based on period)
Deadline Monotonic (when deadline != period)
Preemptive.
Always run current task with shortest period.
Advantages:
Disadvantages:
Example:
Periods:
Schedule:
4
1
3
5
P1
P1
P1
P2
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
P2
P2
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
~ 0.83 for m=2
~ 0.69 for large m
Liu & Layland 1973
Preemptive.
Dynamic Priority
Always run task with earliest deadline relative to
current time.
Example:
Periods:
Schedule:
4
1
3
5
P1
P1
P2
P2
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
P1
P2
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
-
-
-
-
-
-
-
-
-
-
-
-
-
Optimal: Minimizes maximum lateness
Requires known worst-case execution time
Collect process behavior data from real users
Stochastic models of process behavior
Build the thing and do incremental rollout
Collect user feedback and iterate
Very slow to modify things...
Model accuracy is difficult...
Compare algorithm performance on samples
Common research methodology,
but representative sampling is very difficult...
Unschedulable tasks are less predictable
Representing deadlines in hardware non-trivial
(dealing with integer overflow?)
Advantages:
Disadvantages:
where
Guaranteed to hit all deadlines when:
> set of tasks is schedulable
> deadline = start of next period
Linux EDF implementation:
How do you choose a scheduling algorithm?
Need performance data to compare...
Deterministic Modeling:
Queuing Models:
Implementation:
A very good read:
SCHED_DEADLINE
tontinton - Scheduling Internals
Evaluating Scheduling Algorithms
Exercise:
PollEv
A) Yes
B) No