Scheduling Periodic
Scheduling Periodic
T:period D: deadline
Given by the period T The same worst case execution time: C The same relative deadline: D=T (not a restriction) The same relative arrival time: A=0 (not a restriction) The same release time, released as soon as they arrive No sharing resources (consider this later) E.g context switch etc (consider this later)
4
CPU utilization
A task = (C, T)
C/T is the CPU utilization of a task U= Ci/Ti is the CPU utilization of a task set Note that the CPU utilization is a measure on how busy the processor could be during the shortest repeating cycle: T1*T2*...*Tn
U>1 (overload): some task will fail to meet its deadline no matter what algorithms you use! U<=1: it will depend on the scheduling algorithms
If U=1 and the CPU is kept busy (non idle algorithms e.g. EDF), all deadlines will be met
Scheduling Algorithms
Static Cyclic Scheduling (SCS) Earliest Deadline First (EDF) Rate Monotonic Scheduling (RMS) Deadline Monotonic Scheduling (DMS)
Shortest repeating cycle = least common multiple (LCM) Within the cycle, it is possible to construct a static schedule i.e. a time table Schedule task instances according to the time table within each cycle Synchronous programming languages: Esterel, Lustre, Signal
8
The shortest repeating cycle = 80ms All task instances within the cycle:
0 Speed ABS Fuel 20 Speed 40 Speed ABS 60 Speed 80
Speed measurment: C=4ms, T=20ms, D=20ms ABS control: C=10ms,T=40ms, D=40ms Fuel injection: C=40ms,T=80ms, D=80ms Other software with soft deadlines e.g audio, air condition etc
10
80 0
Soft RT tasks speed
4
ABS
14
FUEL-1
64
speed
20
Difficult to modify, e.g adding another task Difficult to handle external events
Huge memory-usage Difficult to construct the time table
A feasible Schedule!
FUEL-3
speed
24
Fuel-2
60
54
ABS
44
speed
40
11
12
Task model
LCM =50ms for tasks with periods: 5ms, 10ms and 25ms LCM =7*13*23=2093 ms for tasks with periods: 7ms, 13ms and 23ms (very much bigger)
a set of independent periodic tasks (not necessarily the simplified task model) Whenever a new task arrive, sort the ready queue so that the task closest to the end of its period assigned the highest priority Preempt the running task if it is not placed in the first of the queue in the last sorting EDF can schedule the task set if any one else can Ci/Ti <= 1 iff the task set is schedulable
14
EDF:
So if possible, manipulate the periods so that they are multiples of each other
Easier to find a feasible schedule and Reduce the size of the static schedule, thus less memory usage
13
Example
EDF: + and
Note that this is just the simple EDF algorithm; it works for all types of tasks: periodic or non periodic
It is simple and works nicely in theory (+) Simple schedulability test: U <= 1 (+) Optimal (+) Best CPU utilization (+)
10
15
35
14
35
Difficult to implement in practice. It is not very often adopted due to the dynamic priority-assignment (expensive to sort the ready queue on-line), which has nothing to do with the periods of tasks. Note that Any task could get the highest priority (-) Non stable: if any task instance fails to meet its deadline, the system is not predictable, any instance of any task may fail (-)
Run-Time Scheduling:
If a task set is schedulable with any fixed-priority scheduling algorithm, it is also schedulable with RMS
17
18
Example
{(20,100),(40,150),(100,350)} T1 20 0 20 100 Pr(T1)=1, Pr(T2)=2, Pr(T3)=3 20 200 20 300
Example
Task set: T1=(2,5), T2=(4,7) U = 2/5 + 4/7= 34/35 ~ 0.97 (schedulable?) RMS priority assignment: Pr(T1)=1, Pr(T2)=2
T2
40
40 150
300
40
0 2
35
T3 0
40
30
10
20 350
19
5 7
35
20
The famous Utilization Bound test (UB test) [by Liu and Layland, 1973: a classic result]
FACT: if U<= n*(21/n-1), then S is schedulable by RMS Note that the bound depends only on the size of the task set
Given a task set S, find X(S) such that U<= X(S) if and only if S is schedulable by RMS (necessary and sufficient test) Note that the bound X(S) for EDF is 1
21
22
Task 2
Task 3
40
100
150
350
0.267
0.286
23
24
Example: RM Scheduling
{(20,100),(40,150),(100,350)} 20 0 20 100 20 200 20 300
40
40 150
300
40
Thus, the test may be too conservative (exact test will be given later)
40
30
10
20 350
25 26
Assume a task set: {(1,3),(1,5),(1,6),(2,10)} CPU utilization U= 1/3+1/5+1/6+2/10=0.899 The utilization bound B(4)=0.756 The task set fails in the UB test due to U>B(4) Question: is the task set schedulable? Answer: YES
{(1,3),(1,5),(1,6),(2,10)}
0 0 0 3 5 6 6 9 10 12 12 15 15
18 20 18
0
27
10
20
28
This is only for the first periods! But we will see that this is enough to tell that the task set is schedullable.
RMS: Summary
What should we do if tasks have the same period? Should we assign the same priority to the tasks? How about the UB test? Is it still sufficient? What happens at run time?
Task model:
Fixed-priority assignment:
Run time scheduling: Preemptive HPF Sufficient schedulability test: U<= n*(21/n-1) Precise/exact schedulability test exists
29
30
RMS: + and
Simple to understand (and remember!) (+) Easy to implement (static/fixed priority assignment)(+) Stable: though some of the lower priority tasks fail to meet deadlines, others may meet deadlines (+) lower CPU utilization (-) Requires D=T (-) Only deal with independent tasks (-) Non-precise schedulability analysis (-) But these are not really disadvantages;they can be fixed (+++)
Note that in our examples, we have assumed that all tasks are released at the same time: this is to consider the critical instant (the worst case senario)
If tasks meet the first deadlines (the first periods), they will do so in the future (why?)
Critical instant of a task is the time at which the release of the task will yield the largest response time. It occurs when the task is released simultaneously with higher priority tasks Note that the start of a task period is not necessarily the same as any of the other periods: but the delay between two releases should be equal to the constant period (otherwise we have jitters)
31
32
Critical instant: the worst case response time for all tasks is given when all tasks are released at the same time Calculate the worst case response time R for each task with deadline D. If R<=D, the task is schedulable/feasible. Repeat the same check for all tasks If all tasks pass the test, the task set is schedulable If some tasks pass the test, they will meet their deadlines even the other dont (stable and predictable) how to calculate the worst case response times?
18 20
Question:
0
33
10
20
34
Let Ri stand for the response time for task i. Then Ri= Ci + j I(i,j)
18 20
W CR=1
Ci is the computing time I(i,j) is the so-called interference of task j to i I(i,j) = 0 if task i has higher priority than j x denotes the least integer larger than x E.g 3.2 = 4, 3 =3, 1.9 =2
WCR=2
18
WCR=3
0
What to do if too many?
10
20
WCR=9
35 36
Ri/Tj is the number of instances of task j during Rj Ri/Tj*Cj is the time needed to execute all instances of task j released within Rj j HP(i) Ri/Tj*Cj is the time needed to execute instances of tasks with higher priorities than task i, released during Rj Rj is the sum of the time required for executing task instances with higher priorities than task j and its own computing time
We need to solve the equation: Ri= Ci + j HP(i) Ri/Tj*Cj This can be done by numerical methods to compute the fixed point of the equation e.g. By iteration: let
Ri0 = Ci + j HP(i) Cj = C1+C2+...+Ci (the first guess) Rik+1 = Ci + j HP(i) Rik/Tj*Cj (the (k+1)th guess) Rim+1>Ti or non schedulable Rim<Ti and Rim+1 = Rim schedulable
37
Example
Assume a task set: {(1,3),(1,5),(1,6),(2,10)} Question: is the task set schedulable? Answer: YES Because R11 = R10 = C1=1 (done) R20 = C2 + C1=2, R21 = C2 + R20/T1*C1=1+ 2/3*1=2 (done)
Order tasks according to their priorities (periods) Use UB test as far as you can until you find the first non-schedulable task Calculate response time for the task and all the tasks with lower priority
39
40
Example
Consider the same task set: {(1,3),(1,5),(1,6),(3,10)} CPU utilization U= 1/3+1/5+1/6+3/10=0.899> B(4)= 0.756
Total utilization: U=0.4+0.267+0.286= 0.953>B(3)=0.779! UB test is inclusive: we need Precise test but we do have U(T1)+U(T2)= 0.4+0.267= 0.667<U(2)=0.828 so we need to calculate R3 only!
41 42
R30 =
C1+C2+C3= 180 R31 = C3+R30/T1*C1+R30/T2*C2 =100+ 180/100*40+180/150*40 =100+2*40+2*40=260 R32 =C3+R31/T1*C1+R31/T2*C2 =100+ 260/100*40+260/150*40=300 R33 =C3+R32/T1*C1+R32/T2*C2 =100+ 300/100*40+300/150*40=300 (done)
Could we calculate the response times by the same equation for different priority assignment?
43
44
Precedence constraints
How to handle precedence constraints?
We can always try the old method: static cyclic scheduling! Alternatively, take the precedence constraints (DAG) into account in priority assignment: the priority-ordering must satisfy the precedence constraints
UB test (simple but conservative) Response time calculation (precise test) Construct a schedule for the first periods
Precise schedulability test is valid: use the same method as beforee to calculate the response times.
assume the first instances arrive at time 0 (critical instant) draw the schedule for the first periods if all tasks are finished before the end of the first periods, schedulable, otherwise NO
45
46
Deadline <= Period Interrupt handling Non zero OH for context switch Non preemptive sections Resource Sharing
RMS is no longer optimal (example?) Utilization bound test must be modified Response time test is still applicable
Assuming that fixed-priority assignment is adopted But considering the critical instant and checking the first deadlines principle are still applicable
47
48
Example
C Task 1 Task 2 Task 3 Task 4 1 1 2 1 T 4 5 6 D 3 5 4
5 6 10 12 11 R1=1 R2=4 R3=3 R4=10 15 4 8 12 16
Task model: the same as for RMS but Di<=Ti Priority-Assignment: tasks with shorter deadline are assigned higher priorities Run-time scheduling: preemptive HPF FACTS:
11 10
DMS is often refered as Rate Monotonic Scheduling for historical reasons and they are so similar
49
50
UB test (sufficient):
Ci/Di <= n*(21/n-1) implies schedulable by DMS
UB test (sufficient, inconclusive) Response time calculation Draw the schedule for the first periods
51
52
You can always use EDF and it is always optimal to schedule tasks with deadlines
We have a precise UB test for EDF for tasks with Di=Ti: U<=1 iff task set is schedulable Unfortunately, for tasks with Di<=Ti, schedulability analysis is more complicated (out of scope of the course, further reading [Giorgio Buttazzos book])
Dynamic priority
53
EDF ?
54
Assume that
Cl is the extra time required to load the context for a new task (load contents of registers etc from TCB) Cs is the extra time required to save the context for a current task (save contents of registers etc to TCB) Note that in most cases, Cl=Cs, which is a parameter depending on hardware Cl Cs Task 1 Task 2
55
Thus, the real computing time for a task should be Ci= Ci+Cl+Cs The schedulability analysis techniques we studied so far are applicable if we use the new computing time C.
Dispatch/context switch
56
60
100
This is wrong!
Released here
IH, task 0
60 10 40
Task 1 Task 2
50 60
Task 2
This is right
57
58
Whenever possible: move code from the interrupt handler to a special application task with the same rate as the interrupt handler to make the interrupt handler (with high priority) as shorter as possible Interrupt processing can be inconsistent with RM priority assignment, and therefore can effect schedulability of task set (previous example)
Interrupt handler runs with high priority despites its period Interrupt processing may delay tasks with shorter periods (deadlines) how to calculate the worst case response time ?
59
60
10
So far, we have assumed that all tasks are preemptive regions of code. This not always the case e.g code for context switch though it may be short, and the short part of the interrupt handler as we considered before
blocked 20 20
100 150
0 0
Task 3
Task 4
60
40
200
350
0
20
20
0
In general, we may assume an extra parameter B in the task model, which is the computing time for the non preemtive section of a task.
60
100
200
150 20
Non preemptive/non interruptible section of 20
62
61
The equation for response time calculation: Ri= Bi + Ci + j HP(i) Ri/Tj*Cj Where Bi is the longest time that task i can be blocked by lower-priority tasks with non preemptive section
Note that a task preempts only one task with lower priority within each period
63
64
Jitter: Example
{(20,100),(40,150),(20, T3)}
T1 20 0 40 0 T3 0 20 100 40 150 20 200 20 300 40 300
So far, we have assumed that tasks are released at a constant rate (at the start of a constant period) This is true in practice and a realistic assumption However, there are situations where the period or rather the release time may jitter or change a little, but the jitter is bounded with some constant J The jitter may cause some task missing deadline
T2
130
20
170
150
300
T3 is activated by T2 when it finishes within each period Note that because the response time for T2 is not a constant, the period between two instances of T3 is not a constant: 170, 130
65 66
11
Jitter: Definition
Jitter: Example
{(20,100),(40,150),(20, T3)}
T1 T2 0 T3 0 20 0 40 20 100 40 150
J(biggest)=maximal delay from period-start J(smallest)=minimal delay from period-start Jitter= J(biggest)-J(smallest) Jitter = the maximal length of the interval in which a task may be released non-deterministically If J(biggest)=J(smallest), then NO JITTER and therefore no influence on the other tasks with lower priorities
67
130
20
170
150
300
T3 is activated by T2 by the end of each instance J(biggest)= R2(worst case), J(smallest)= R2(best case) Jitter = J(biggest)- J(smallest)=60-40=20
68
Jitter: Example
{(20,100),(40,150),(20, T3)}
T1
T2 0 T3 0 20 0 40 20 100 40 150 20 200 20 300 40
Rlow
One release Task L 0
90
20
210
300
Tlow Jhigh
One more release due to the jitter W hich preempts L, one more time
150
300
Task H 0
T3 is activated by T2 at any time during its execution of an instance J(biggest)= R2(worst case), J(smallest)= R2(best case)-C2 Jitter = J(biggest)- J(smallest)=60-0=60
69
Thigh
70
Rlow
One release Task L 0
Tlow
Jhigh
Task H 0
One more release due to the jitter Which preempts L, one more time
Thigh
2Thigh
Task L will be preempted at least 2 times if Rlow > Thigh -Jhigh Task L will be preempted at least 3 times if Rlow > 2 *T high -Jhigh ... Task L will be preempted at least n times if Rlow > (n-1)* Thigh Jhigh Thus (Rlow +J high)/Tj > n-1 the largest n satisfying the condition is given by n= (Rlow + Jhigh)/ Thigh
71
72
12
Ri* = Ri + Ji(biggest)
Ri= Ci + j HP(i)
(Ri+Jj)/Tj*C
73
74
Let
CS(k,S) denote the computing time for the critical section that task k uses semaphore S. Use(S) is the set of tasks using S
Then for HLP and PCP, the maximal blocking time RSi and response time Ri for task i is as follows:
75
Summary: + and
Simple, and reliable, may be difficult to construct the time table and difficult to modify and (inflexible)
Simple in theory, but difficult to implement, non-stable no precise analysis for tasks D<T
Simple in theory and practice, and easy to implement Similar to RMS
77
78
13