Chapter 4

Integre Technical Publishing Co., Inc.
C H A P T E R
Liu
January 13, 2000 8:46 a.m.
chap4
page 60
Commonly Used Approaches

to Real-Time Scheduling
This chapter provides a brief overview of three commonly used approaches to scheduling realtime systems: clock-driven, weighted round-robin and priority-driven. The subsequent five
chapters will study in depth the clock-driven and priority-driven approaches, with particular
emphasis on the latter. We need to be aware of several important facts about priority-driven
scheduling. They are presented here and shown to be true, so they will not be buried in the
details later. We will use these facts to motivate and support the arguments presented in later
chapters. The weighted round-robin approach is used primarily for scheduling real-time traffic
in high-speed switched networks. It will be described in depth in Chapter 11. We discuss here
why this approach is not ideal for scheduling jobs on CPUs.
4.1
CLOCK-DRIVEN APPROACH
As the name implies, when scheduling is clock-driven (also called time-driven), decisions on
what jobs execute at what times are made at specific time instants. These instants are chosen
a priori before the system begins execution. Typically, in a system that uses clock-driven
scheduling, all the parameters of hard real-time jobs are fixed and known. A schedule of the
jobs is computed off-line and is stored for use at run time. The scheduler schedules the jobs
according to this schedule at each scheduling decision time. In this way, scheduling overhead
during run-time can be minimized.
A frequently adopted choice is to make scheduling decisions at regularly spaced time
instants. One way to implement a scheduler that makes scheduling decisions periodically is
to use a hardware timer. The timer is set to expire periodically without the intervention of
the scheduler. When the system is initialized, the scheduler selects and schedules the job(s)
that will execute until the next scheduling decision time and then blocks itself waiting for
the expiration of the timer. When the timer expires, the scheduler awakes and repeats these
actions.
60
Liu
January 13, 2000 8:46 a.m.
Section 4.2
4.2
chap4
page 61
Weighted Round-Robin Approach
61
WEIGHTED ROUND-ROBIN APPROACH

The round-robin approach is commonly used for scheduling time-shared applications. When
jobs are scheduled on a round-robin basis, every job joins a First-in-first-out (FIFO) queue
when it becomes ready for execution. The job at the head of the queue executes for at most
one time slice. (A time slice is the basic granule of time that is allocated to jobs. In a timeshared environment, a time slice is typically in the order of tens of milliseconds.) If the job
does not complete by the end of the time slice, it is preempted and placed at the end of the
queue to wait for its next turn. When there are n ready jobs in the queue, each job gets one time
slice every n time slices, that is, every round. Because the length of the time slice is relatively
short, the execution of every job begins almost immediately after it becomes ready. In essence,
each job gets 1/nth share of the processor when there are n jobs ready for execution. This is
why the round-robin algorithm is also called the processor-sharing algorithm.
The weighted round-robin algorithm has been used for scheduling real-time traffic in
high-speed switched networks. It builds on the basic round-robin scheme. Rather than giving
all the ready jobs equal shares of the processor, different jobs may be given different weights.
Here, the weight of a job refers to the fraction of processor time allocated to the job. Specifically, a job with weight wt gets wt time slices every round, and the length of a round is equal
to the sum of the weights of all the ready jobs. By adjusting the weights of jobs, we can speed
up or retard the progress of each job toward its completion.
By giving each job a fraction of the processor, a round-robin scheduler delays the completion of every job. If it is used to schedule precedence constrained jobs, the response time
of a chain of jobs can be unduly large. For this reason, the weighted round-robin approach
is not suitable for scheduling such jobs. On the other hand, a successor job may be able to
incrementally consume what is produced by a predecessor (e.g., as in the case of a UNIX
pipe). In this case, weighted round-robin scheduling is a reasonable approach, since a job and
its successors can execute concurrently in a pipelined fashion. As an example, we consider
the two sets of jobs, J1 = {J1,1 , J1,2 } and J2 = {J2,1 , J2,2 }, shown in Figure 41. The release
times of all jobs are 0, and their execution times are 1. J1,1 and J2,1 execute on processor
P1 , and J1,2 and J2,2 execute on processor P2 . Suppose that J1,1 is the predecessor of J1,2 ,
and J2,1 is the predecessor of J2,2 . Figure 41(a) shows that both sets of jobs (i.e., the second
jobs J1,2 and J2,2 in the sets) complete approximately at time 4 if the jobs are scheduled in
a weighted round-robin manner. (We get this completion time when the length of the time
slice is small compared with 1 and the jobs have the same weight.) In contrast, the schedule in
Figure 41(b) shows that if the jobs on each processor are executed one after the other, one of
the chains can complete at time 2, while the other can complete at time 3. On the other hand,
suppose that the result of the first job in each set is piped to the second job in the set. The
latter can execute after each one or a few time slices of the former complete. Then it is better
to schedule the jobs on the round-robin basis because both sets can complete a few time slices
after time 2.
Indeed, the transmission of each message is carried out by switches en route in a pipeline
fashion. A switch downstream can begin to transmit an earlier portion of the message as soon
as it receives the portion without having to wait for the arrival of the later portion of the
message. The weighted round-robin approach does not require a sorted priority queue, only
a round-robin queue. This is a distinct advantage for scheduling message transmissions in
ultrahigh-speed networks, since priority queues with the required speed are expensive. In
62
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 62
Commonly Used Approaches to Real-Time Scheduling
J1,1
J1,2
J2,1
J2,2
P1
J1,1 & J2,1
P2
(a)
J1,2 & J2,2

0
P1
J1,1
J1,2
FIGURE 41
J2,1
P2
0
(b)
J2,2
2
Example illustrating round-robin scheduling of precedence-constrained jobs.
Chapter 11 we will describe in detail the application of weighted round-robin algorithms to

network scheduling.
4.3
PRIORITY-DRIVEN APPROACH
The term priority-driven algorithms refers to a large class of scheduling algorithms that never
leave any resource idle intentionally. Stated in another way, a resource idles only when no job
requiring the resource is ready for execution. Scheduling decisions are made when events such
as releases and completions of jobs occur. Hence, priority-driven algorithms are event-driven.
Other commonly used names for this approach are greedy scheduling, list scheduling
and work-conserving scheduling. A priority-driven algorithm is greedy because it tries to
make locally optimal decisions. Leaving a resource idle while some job is ready to use the
resource is not locally optimal. So when a processor or resource is available and some job can
use it to make progress, such an algorithm never makes the job wait. We will return shortly to
illustrate that greed does not always pay; sometimes it is better to have some jobs wait even
when they are ready to execute and the resources they require are available.
The term list scheduling is also descriptive because any priority-driven algorithm can
be implemented by assigning priorities to jobs. Jobs ready for execution are placed in one or
Liu
January 13, 2000 8:46 a.m.
Section 4.3
chap4
page 63
Priority-Driven Approach
63
more queues ordered by the priorities of the jobs. At any scheduling decision time, the jobs
with the highest priorities are scheduled and executed on the available processors. Hence,
a priority-driven scheduling algorithm is defined to a great extent by the list of priorities it
assigns to jobs; the priority list and other rules, such as whether preemption is allowed, define
the scheduling algorithm completely.
Most scheduling algorithms used in nonreal-time systems are priority-driven. Examples
include the FIFO (First-In-First-Out) and LIFO (Last-In-First-Out) algorithms, which assign
priorities to jobs according their release times, and the SETF (Shortest-Execution-Time-First)
and LETF (Longest-Execution-Time-First) algorithms, which assign priorities on the basis of
job execution times. Because we can dynamically change the priorities of jobs, even roundrobin scheduling can be thought of as priority-driven: The priority of the executing job is
lowered to the minimum among all jobs waiting for execution after the job has executed for a
time slice.
Figure 42 gives an example. The task graph shown here is a classical precedence graph;
all its edges represent precedence constraints. The number next to the name of each job is its
execution time. J5 is released at time 4. All the other jobs are released at time 0. We want to
schedule and execute the jobs on two processors P1 and P2 . They communicate via a shared
memory. Hence the costs of communication among jobs are negligible no matter where they
are executed. The schedulers of the processors keep one common priority queue of ready jobs.
The priority list is given next to the graph: Ji has a higher priority than Jk if i < k. All the
jobs are preemptable; scheduling decisions are made whenever some job becomes ready for
execution or some job completes.
Figure 42(a) shows the schedule of the jobs on the two processors generated by the
priority-driven algorithm following this priority assignment. At time 0, jobs J1 , J2 , and J7
are ready for execution. They are the only jobs in the common priority queue at this time.
Since J1 and J2 have higher priorities than J7 , they are ahead of J7 in the queue and hence are
scheduled. The processors continue to execute the jobs scheduled on them except when the
following events occur and new scheduling decisions are made.
At time 1, J2 completes and, hence, J3 becomes ready. J3 is placed in the priority queue
ahead of J7 and is scheduled on P2 , the processor freed by J2 .
At time 3, both J1 and J3 complete. J5 is still not released. J4 and J7 are scheduled.
At time 4, J5 is released. Now there are three ready jobs. J7 has the lowest priority
among them. Consequently, it is preempted. J4 and J5 have the processors.
At time 5, J4 completes. J7 resumes on processor P1 .
At time 6, J5 completes. Because J7 is not yet completed, both J6 and J8 are not ready
for execution. Consequently, processor P2 becomes idle.
J7 finally completes at time 8. J6 and J8 can now be scheduled and they are.
Figure 42(b) shows a nonpreemptive schedule according to the same priority assignment. Before time 4, this schedule is the same as the preemptive schedule. However, at time
4 when J5 is released, both processors are busy. It has to wait until J4 completes (at time 5)
before it can begin execution. It turns out that for this system this postponement of the higher
priority job benefits the set of jobs as a whole. The entire set completes 1 unit of time earlier
according to the nonpreemptive schedule.
64
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 64
J1, 3
J2, 1
J3, 2
J5, 2
J6, 4
J7, 4
J8, 1
J1
P1
0
J4, 2
(J1, J2,, J8)
J4
2
J7
5
J6
7
10
11
12
(a)
P2 J2
J3
P1
J7
J1
J5
4
J4
4
J8
6
J5
5
10
11
12
10
11
12
J6
9
(b)
P2 J2
0
J3
1
J7
3
J8
6
10
11
12
FIGURE 42 Example of priority-driven scheduling. (a) Preemptive (b) Nonpreemptive.
In general, however, nonpreemptive scheduling is not better than preemptive scheduling.

A fundamental question is, when is preemptive scheduling better than nonpreemptive scheduling and vice versa? It would be good if we had some rule with which we could determine from
the given parameters of the jobs whether to schedule them preemptively or nonpreemptively.
Unfortunately, there is no known answer to this question in general. In the special case when
jobs have the same release time, preemptive scheduling is better when the cost of preemption
is ignored. Specifically, in a multiprocessor system, the minimum makespan (i.e., the response
time of the job that completes last among all jobs) achievable by an optimal preemptive algorithm is shorter than the makespan achievable by an optimal nonpreemptive algorithm. A
natural question here is whether the difference in the minimum makespans achievable by the
two classes of algorithms is significant, in particular, whether the theoretical gain in makespan
achievable by preemption is enough to compensate for the context switch overhead of preemption. The answer to this question is only known for the two-processor case. Coffman and
Liu
January 13, 2000 8:46 a.m.
Section 4.5
chap4
page 65
Effective Release Times and Deadlines
65
Garey [CoGa] recently proved that when there are two processors, the minimum makespan
achievable by nonpreemptive algorithms is never more than 4/3 times the minimum makespan
achievable by preemptive algorithms when the cost of preemption is negligible. The proof of
this seemingly simple results is too lengthy to be included here.
4.4
DYNAMIC VERSUS STATIC SYSTEMS

In the above example, jobs that are ready for execution are placed in a priority queue common
to all processors. When a processor is available, the job at the head of the queue executes on
the processor. We will refer to such a multiprocessor system as a dynamic system, because
jobs are dynamically dispatched to processors. In the example in Figure 42, we allowed each
preempted job to resume on any processor and hence, jobs are migratable. We say that a job
migrates if it starts execution on a processor, is preempted, and later resumes on a different
processor.
Another approach to scheduling in multiprocessor and distributed systems is to partition
the jobs in the system into subsystems and assign and bind the subsystems statically to the
processors. Jobs are moved among processors only when the system must be reconfigured,
that is, when the operation mode of the system changes or some processor fails. Such a system is called a static system, because the system is statically configured. If jobs on different
processors are dependent, the schedulers on the processors must synchronize the jobs according to some synchronization and resource access-control protocol. Except for the constraints
thus imposed, the jobs on each processor are scheduled by themselves.
As an example, a partition and assignment of the jobs in Figure 42 put J1 , J2 , J3 ,
and J4 on P1 and the remaining jobs on P2 . The priority list is segmented into two parts:
(J1 , J2 , J3 , J4 ) and (J5 , J6 , J7 , J8 ). The scheduler of processor P1 uses the former while the
scheduler of processor P2 uses the latter. It is easy to see that the jobs on P1 complete by time
8, and the jobs on P2 complete by time 11. Moreover, J2 completes by time 4 while J6 starts
at time 6. Therefore, the precedence constraint between them is satisfied.
In this example, the response of the static system is just as good as that of the dynamic
system. Intuitively, we expect that we can get better average responses by dynamically dispatching and executing jobs. In later chapters we will return to this discussion. Specifically,
we will demonstrate that while dynamic systems may be more responsive on the average,
their worst-case real-time performance may be poorer than static systems. More importantly,
we do not yet have reliable techniques to validate the timing constraints of dynamic systems
while such techniques exist for static systems. For this reason, most hard real-time systems
built today are static.
4.5
EFFECTIVE RELEASE TIMES AND DEADLINES

The given release times and deadlines of jobs are sometimes inconsistent with the precedence
constraints of the jobs. By this, we mean that the release time of a job may be later than that of
its successors, and its deadline may be earlier than that of its predecessors. Therefore, rather
than working with the given release times and deadlines, we first derive a set of effective
66
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 66
release times and deadlines from these timing constraints, together with the given precedence
constraints. The derived timing constraints are consistent with the precedence constraints.
When there is only one processor, we can compute the derived constraints according to
the following rules:
Effective Release Time: The effective release time of a job without predecessors is equal to
its given release time. The effective release time of a job with predecessors is equal to
the maximum value among its given release time and the effective release times of all
of its predecessors.
Effective Deadline: The effective deadline of a job without a successor is equal to its given
deadline. The effective deadline of a job with successors is equal to the minimum value
among its given deadline and the effective deadlines of all of its successors.
The effective release times of all the jobs can be computed in one pass through the precedence
graph in O(n 2 ) time where n is the number of jobs. Similarly, the effective deadlines can be
computed in O(n 2 ) time.
As an example, we look at the set of jobs in Figure 43. The numbers in the parentheses
next to the name of each job are its given release times and deadlines. Because J1 and J2
have no predecessors, their effective release times are the given release times, that is, 2 and
0, respectively. The given release time of J3 is 1, but the latest effective release time of its
predecessors is 2, that of J1 . Hence, the effective release time of J3 is 2. You can repeat this
procedure and find that the effective release times of the rest of the jobs are 4, 2, 4, and 6,
respectively. Similarly, J6 and J7 have no successors, and their effective deadlines are equal
to their given deadlines, 20 and 21, respectively. Since the effective deadlines of the successors
of J4 and J5 are larger than the given deadlines of J4 and J5 , the effective deadlines of J4 and
J5 are equal to their given deadlines. On the other hand, the given deadline of J3 is equal to 12,
which is larger than the minimum value of 8 among the effective deadlines of its successors.
Hence, the effective deadline of J3 is 8. In a similar way, we find that the effective deadlines
of J1 and J2 are 8 and 7, respectively.
You may have noticed that the calculation of effective release times and deadlines does
not take into account the execution times of jobs. More accurately, the effective deadline of
a job should be as early as the deadline of each of its successors minus the execution time of
the successor. The effective release time of a job is that of its predecessor plus the execution
time of the predecessor. The more accurate calculation is unnecessary, however, when there is
only one processor. Gary and Johnson [GaJo77] have shown that it is feasible to schedule any
J1 (2, 10)
J3 (1, 12)
J4 (4, 9)
J5 (1, 8)
J2 (0, 7)
FIGURE 43
Example of effective timing constraints.
J6 (0, 20)
J7 (6, 21)
Liu
Section 4.6
January 13, 2000 8:46 a.m.
chap4
page 67
Optimality of the EDF and LST Algorithms
67
set of jobs on a processor according to their given release times and deadlines if and only if it
is feasible to schedule the set according to their effective release times and deadlines defined
above. When there is only one processor and jobs are preemptable, working with the effective
release times and deadlines allows us to temporarily ignore the precedence constraints and
treat all the jobs as if they are independent. Of course, by doing so, it is possible for an
algorithm to produce an invalid schedule that does not meet some precedence constraint.
For example, J1 and J3 in Figure 43 have the same effective release time and deadline.
An algorithm which ignores the precedence constraint between them may schedule J3 in an
earlier interval and J1 in a later interval. If this happens, we can always add a step to swap the
two jobs, that is, move J1 to where J3 is scheduled and vice versa. This swapping is always
possible, and it transforms an invalid schedule into a valid one.
Hereafter, by release times and deadlines, we will always mean effective release times
and deadlines. When there is only one processor and jobs are preemptable, we will ignore the
precedence constraints.
4.6
OPTIMALITY OF THE EDF AND LST ALGORITHMS

A way to assign priorities to jobs is on the basis of their deadlines. In particular, the earlier
the deadline, the higher the priority. The priority-driven scheduling algorithm based on this
priority assignment is called the Earliest-Deadline-First (EDF) algorithm. This algorithm is
important because it is optimal when used to schedule jobs on a processor as long as preemption is allowed and jobs do not contend for resources. This fact is stated formally below.
T HEOREM 4.1. When preemption is allowed and jobs do not contend for resources,
the EDF algorithm can produce a feasible schedule of a set J of jobs with arbitrary
release times and deadlines on a processor if and only if J has feasible schedules.
Proof. The proof is based on the following fact: Any feasible schedule of J can be
systematically transformed into an EDF schedule (i.e., a schedule produced by the EDF
algorithm). To see why, suppose that in a schedule, parts of Ji and Jk are scheduled
in intervals I1 and I2 , respectively. Furthermore, the deadline di of Ji is later than the
deadline dk of Jk , but I1 is earlier than I2 as shown in Figure 44.
There are two cases. In the first case, the release time of Jk may be later than
the end of I1 . Jk cannot be scheduled in I1 ; the two jobs are already scheduled on the
EDF basis in these intervals. Hence, we need to consider only the second case where
the release time rk of Jk is before the end of I1 ; without loss of generality, we assume
that rk is no later than the beginning of I1 .
To transform the given schedule, we swap Ji and Jk . Specifically, if the interval
I1 is shorter than I2 , as shown in Figure 44, we move the portion of Jk that fits in I1
forward to I1 and move the entire portion of Ji scheduled in I1 backward to I2 and place
it after Jk . The result is as shown in Figure 44(b). Clearly, this swap is always possible.
We can do a similar swap if the interval I1 is longer than I2 : We move the entire portion
of Jk scheduled in I2 to I1 and place it before Ji and move the portion of Ji that fits in
I2 to the interval. The result of this swap is that these two jobs are now scheduled on the
68
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 68
I1
I2
Ji
Jk
(a)
rk
dk
Jk
Jk
FIGURE 44
Ji
Jk
Jk
dk
Ji
(b)
(c)
Transformation of a non-EDF schedule into an EDF schedule.
EDF basis. We repeat this transformation for every pair of jobs that are not scheduled
on the EDF basis according to the given non-EDF schedule until no such pair exists.
The schedule obtained after this transformation may still not be an EDF schedule
if some interval is left idle while there are jobs ready for execution but are scheduled in
a later interval (e.g., as in the schedule in Figure 44(b).) We can eliminate such an idle
interval by moving one or more of these jobs forward into the idle interval and leave the
interval where the jobs were scheduled idle. This is clearly always possible. We repeat
this process if necessary until the processor never idles when there are jobs ready for
execution as in Figure 44(c).
That the preemptive EDF algorithm can always produce a feasible schedule as
long as feasible schedules exist follows straightforwardly from the fact that every
feasible schedule can be transformed into a preemptive EDF schedule. If the EDF algorithm fails to produce a feasible schedule, then no feasible schedule exists. (If a feasible
schedule were to exist, it could be transformed into an EDF schedule, which contradicts
the statement that the EDF algorithm fails to produce a feasible schedule.)
When the goal of scheduling is to meet deadlines, there is no advantage to completing
any job sooner than necessary. We may want to postpone the execution of hard real-time
jobs for some reason (e.g., to enable soft real-time jobs, whose response times are important,
to complete earlier). For this reason, we sometimes also use the latest release time (LRT)
algorithm (or reverse EDF algorithm). This algorithm treats release times as deadlines and
deadlines as release times and schedules the jobs backwards, starting from the latest deadline
of all jobs, in priority-driven manner, to the current time. In particular, the priorities are
based on the release times of jobs: the later the release time, the higher the priority. Because
it may leave the processor idle when there are jobs ready for execution, the LRT algorithm is
not a priority-driven algorithm.
Liu
Section 4.6
January 13, 2000 8:46 a.m.
chap4
page 69
Optimality of the EDF and LST Algorithms
J1, 3 (0, 6]
69
J2, 2 (5, 8]
J3, 2 (2, 7]
J3
J1
0
2
FIGURE 45
J2
6
Example illustrating the LRT algorithm.
Figure 45 gives an example. In the precedence graph, the number next to the job name
is the execution time of the job. Its feasible interval is given by the range of time next to
its execution time. The latest deadline among all jobs is 8. Hence time starts at 8 and goes
backwards to 0. At time 8, J2 is ready and is scheduled. At time 7, J3 is also ready to be
scheduled, but because J2 has a later release time, it has a higher priority. Consequently, J2
is scheduled from 7 to 6. When J2 completes at time 6, J1 is ready. However, J3 has a
higher priority and is, therefore, scheduled from 6 to 4. Finally J1 is scheduled from 4 to 1.
The result is a feasible schedule.
The following corollary states that the LRT algorithm is also optimal under the same
conditions that the EDF algorithm is optimal. Its proof follows straightforwardly from the
proof of Theorem 4.1.
C OROLLARY 4.2. When preemption is allowed and jobs do not contend for resources, the LRT algorithm can produce a feasible schedule of a set J of jobs with
arbitrary release times and deadlines on a processor if and only if feasible schedules of
J exist.
Another algorithm that is optimal for scheduling preemptive jobs on one processor is
the Least-Slack-Time-First (LST) algorithm (also called the Minimum-Laxity-First (MLF) algorithm) [LeWh, Mok]. At any time t, the slack (or laxity) of a job with deadline at d is equal
to d t minus the time required to complete the remaining portion of the job. Take the job
J1 in Figure 45 as an example. It is released at time 0, its deadline is 6, and its execution
time is 3. Hence, its slack is equal to 3 at time 0. The job starts to execute at time 0. As long
as it executes, its slack remains at 3, because at any time t before its completion, its slack is
6 t (3 t). Now suppose that it is preempted at time 2 by J3 , which executes from time
2 to 4. During this interval, the slack of J1 decreases from 3 to 1. (At time 4, the remaining
execution time of J1 is 1, so its slack is 6 4 1 = 1.)
The LST algorithm assigns priorities to jobs based on their slacks: the smaller the slack,
the higher the priority. The following theorem states that the LST algorithm is also optimal.
Its proof is left as an exercise at the end of this chapter.
70
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 70
T HEOREM 4.3. When preemption is allowed and jobs do not contend for resources,
the LST (MLF) algorithm can produce a feasible schedule of a set J of jobs with arbitrary release times and deadlines on a processor if and only if feasible schedules of J
exist.
While the EDF algorithm does not require any knowledge of the execution times of
jobs, the LST algorithm does. This is a serious disadvantage. As discussed earlier, the actual
execution times of jobs are often not known until the jobs complete. Obviously, it is impossible
for us to calculate the actual amounts of slack under this circumstance. We typically calculate
the slack of each job based on its maximum execution time ei + when the range [ei , ei + ] of
execution time ei of every job is relatively small. Furthermore, we require that the maximum
(and sometimes even the actual) execution time of each sporadic or aperiodic job become
known upon its arrival since this knowledge is needed for slack computation.
4.7
NONOPTIMALITY OF THE EDF AND THE LST ALGORITHMS

It is natural to ask here whether the EDF and the LST algorithms remain optimal if preemption
is not allowed or there is more than one processor. Unfortunately, the answer is no.
The fact that the EDF and the LST algorithms are optimal only when preemption is
allowed is illustrated by the example in Figure 46. The system shown in this figure has
three independent, nonpreemptable jobs J1 , J2 , and J3 . Their release times are 0, 2 and 4,
respectively, and are indicated by the arrows above the schedules. Their execution times are
r1
r2
r3
J1
0
J2
2
J3
6
10
12
14
J3 misses its deadline

(a)
J1
0
J3
2
J2
8
10
12
14
(b)
FIGURE 46 Example illustrating nonoptimality of the nonpreemptive EDF algorithm. (a) An EDF schedule. (b)
A non-EDF schedule.
Section 4.7
Liu
January 13, 2000 8:46 a.m.
chap4
page 71
Nonoptimality of the EDF and the LST Algorithms
71
3, 6, and 4; and their deadlines are 10, 14, and 12, respectively. Figure 46(a) shows the
schedule produced by the EDF algorithm. In particular, when J1 completes at time 3, J2 has
already been released but not J3 . Hence, J2 is scheduled. When J3 is released at time 4, J2 is
executing. Even though J3 has an earlier deadline and, hence, a higher priority, it must wait
until J2 completes because preemption is not allowed. As a result, J3 misses its deadline. It is
easy to see that the LST algorithm would produce the same infeasible schedule. The fact that
these three jobs can meet their deadlines is demonstrated by the feasible schedule in Figure
46(b). At time 3 when J1 completes, the processor is left idle, even though J2 is ready for
execution. Consequently, when J3 is released at 4, it can be scheduled ahead of J2 , allowing
both jobs to meet their deadlines.
We note that the schedule in Figure 46(b) cannot be produced by any priority-driven
scheduling algorithm. By definition, a priority-driven algorithm never leaves a processor idle
when there are jobs ready to use the processor. This example illustrates the fact that not only
nonpreemptive EDF and LST algorithms are not optimal, but also no nonpreemptive prioritydriven algorithm is optimal when jobs have arbitrary release times, execution times, and deadlines.
The example in Figure 47 shows that the EDF algorithm is not optimal for scheduling
preemptable jobs on more than one processor. The system in this figure also contains three
jobs, J1 , J2 , and J3 . Their execution times are 1, 1, and 5 and their deadlines are 1, 2, and
5, respectively. The release times of all three jobs are 0. The system has two processors.
According to the EDF algorithm, J1 and J2 are scheduled on the processors at time 0 because
they have higher priorities. The result is the schedule in Figure 47(a), and J3 misses its
deadline.
On the other hand, an algorithm which assigns a higher priority to J3 in this case can
feasibly schedule the jobs. An example of such algorithms is the LST algorithm. The slacks
of the J1 , J2 , and J3 in Figure 47 are 0, 1, and 0, respectively. Hence, this algorithm would
P1
J1
J3
P2
J2
P1
J1
0
P2
J3 misses its deadline
J2
1
(b)
J3
0
(a)
FIGURE 47 Example illustrating nonoptimality of the EDF algorithm for multiprocessor scheduling. (a) The EDF
schedule. (b) A feasible schedule.
72
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 72
produce the feasible schedule in Figure 47(b). Unfortunately, the LST algorithm is also not
optimal for scheduling jobs on more than one processor, as demonstrated by the example in
Problem 4.4.
4.8
CHALLENGES IN VALIDATING TIMING CONSTRAINTS IN PRIORITY-DRIVEN SYSTEMS

Compared with the clock-driven approach, the priority-driven scheduling approach has many
advantages. As examples, you may have noticed that priority-driven schedulers are easy to implement. Many well-known priority-driven algorithms use very simple priority assignments,
and for these algorithms, the run-time overhead due to maintaining a priority queue of ready
jobs can be made very small. A clock-driven scheduler requires the information on the release times and execution times of the jobs a priori in order to decide when to schedule them.
In contrast, a priority-driven scheduler does not require most of this information, making it
much better suited for applications with varying time and resource requirements. You will
see in later chapters other advantages of the priority-driven approach which are at least as
compelling as these two.
Despite its merits, the priority-driven approach has not been widely used in hard realtime systems, especially safety-critical systems, until recently. The major reason is that the
timing behavior of a priority-driven system is nondeterministic when job parameters vary.
Consequently, it is difficult to validate that the deadlines of all jobs scheduled in a prioritydriven manner indeed meet their deadlines when the job parameters vary. In general, this
validation problem [LiHa] can be stated as follows: Given a set of jobs, the set of resources
available to the jobs, and the scheduling (and resource access-control) algorithm to allocate
processors and resources to jobs, determine whether all the jobs meet their deadlines.
4.8.1
Anomalous Behavior of Priority-Driven Systems

Figure 48 gives an example illustrating why the validation problem is difficult when the
scheduling algorithm is priority-driven and job parameters may vary. The simple system contains four independent jobs. The jobs are scheduled on two identical processors in a prioritydriven manner. The processors maintain a common priority queue, and the priority order is
J1 , J2 , J3 , and J4 with J1 having the highest priority. In other words, the system is dynamic.
The jobs may be preempted but never migrated, meaning that once a job begins execution on
a processor, it is constrained to execute on that processor until completion. (Many systems fit
this simple model. For example, the processors may model two redundant data links connecting a source and destination pair, and the jobs are message transmissions over the links. The
processors may also model two replicated database servers, and the jobs are queries dynamically dispatched by a communication processor to the database servers. The release times,
deadlines, and execution times of the jobs are listed in the table.) The execution times of all
the jobs are fixed and known, except for J2 . Its execution time can be any value in the range
[2, 6].
Suppose that we want to determine whether the system meets all the deadlines and
whether the completion-time jitter of every job (i.e., the difference between the latest and the
earliest completion times of the job) is no more than 4. A brute force way to do so is to simulate the system. Suppose that we schedule the jobs according their given priorities, assuming
Section 4.8
Liu
January 13, 2000 8:46 a.m.
ri
di
[ei-, ei+]
J1
10
J2
10
[2, 6]
J3
15
J4
20
10
P1
J1
P2
J2
10
15
20
(a)
J4
16
J1
P1
J2
J4
J3
(b)
J4
P2
73
J3
P1
page 73
Challenges in Validating Timing Constraints in Priority-Driven Systems
P2
chap4
20
J1
J2
J4
J3
P1
J1
P2
J2
(c)
J4
21
J3
(d)
J4
5
15
FIGURE 48 Example illustrating scheduling anomalies.
74
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 74
first that the execution time of J2 has the maximum value 6 and then that it has the minimum
value 2. The resultant schedules are shown in Figure 48(a) and (b), respectively. Looking
at these schedules, we might conclude that all jobs meet their deadlines, and the completiontime jitters are sufficiently small. This would be an incorrect conclusion, as demonstrated by
the schedules in Figure 48(c) and (d). As far as J4 is concerned, the worst-case schedule
is the one shown in Figure 48(c); it occurs when the execution time of J2 is 3. According
to this schedule, the completion time of J4 is 21; the job misses its deadline. The best-case
schedule for J4 is shown in Figure 48(d); it occurs when the execution time of J2 is 5. From
this schedule, we see that J4 can complete as early as time 15; its completion-time jitter exceeds the upper limit of 4. To find the worst-case and best-case schedules, we must try all the
possible values of e2 .
The phenomenon illustrated by this example is known as a scheduling anomaly, an
unexpected timing behavior of priority-driven systems. Graham [Grah] has shown that the
completion time of a set of nonpreemptive jobs with identical release times can be later when
more processors are used to execute them and when they have shorter execution times and
fewer dependencies. (Problem 4.5 gives the well-known illustrative example.) Indeed, when
jobs are nonpreemptable, scheduling anomalies can occur even when there is only one processor. For example, suppose that the execution time e1 of the job J1 in Figure 46 can be either
3 or 4. The figure shows that J3 misses its deadline when e1 is 3. However, J3 would complete
at time 8 and meet its deadline if e1 were 4. We will see in later chapters that when jobs have
arbitrary release times and share resources, scheduling anomalies can occur even when there
is only one processor and the jobs are preemptable.
Scheduling anomalies make the problem of validating a priority-driven system difficult
whenever job parameters may vary. Unfortunately, variations in execution times and release
times are often unavoidable. If the maximum range of execution times of all n jobs in a system
is X , the time required to find the latest and earliest completion times of all jobs is O(X n ) if
we were to find these extrema by exhaustive simulation or testing. Clearly, such a strategy is
impractical for all but the smallest systems of practical interest.
4.8.2
Predictability of Executions
When the timing constraints are specified in terms of deadlines of jobs, the validation problem
is the same as that of finding the worst-case (the largest) completion time of every job. This
problem is easy whenever the execution behavior of the set J is predictable, that is, whenever
the system does not have scheduling anomalies. To define predictability more formally, we
call the schedule of J produced by the given scheduling algorithm when the execution time
of every job has its maximum value the maximal schedule of J. Similarly, the schedule of
J produced by the given scheduling algorithm when the execution time of every job has its
minimum value is the minimal schedule. When the execution time of every job has its actual
value, the resultant schedule is the actual schedule of J. So, the schedules in Figure 48(a)
and (b) are the maximal and minimal schedules, respectively, of the jobs in that system, and
all the schedules shown in the figure are possible actual schedules.
Since the range of execution time of every job is known, the maximal and minimal
schedules of J can easily be constructed when the release-time jitters are negligible. (We
assume that release times of all jobs are fixed and known for this discussion. How releasetime jitters further complicate the validation problem is beyond the scope of our discussion
Section 4.8
Liu
January 13, 2000 8:46 a.m.
chap4
page 75
Challenges in Validating Timing Constraints in Priority-Driven Systems
75
here.) In contrast, its actual schedule is unknown because the actual values of the execution
times are unknown.
The predictability of job execution (behavior) is an intuitive notion: The execution of
J under the given priority-driven scheduling algorithm is predictable if the actual start time
and actual completion time of every job according to the actual schedule are bounded by
its start times and completion times according to the maximal and minimal schedules. More
formally, we let s(Ji ) be the instant of time at which the execution of Ji begins according to
the actual schedule of J. s(Ji ) is the (actual) start time of Ji . Let s + (Ji ) and s (Ji ) be the
start times of Ji according to the maximal schedule and minimal schedule of J, respectively.
These start times can easily be found by constructing the maximal and minimal schedules and
observing when Ji starts according to these schedules. We say that Ji is start-time predictable
if s (Ji ) s(Ji ) s + (Ji ). As an example, for the job J4 in Figure 48, s (J4 ) is 2. s + (J4 )
is 6. Its actual start time is in the range [2, 6]. Therefore, J4 is start-time predictable.
Similarly, let f (Ji ) be the actual completion time (also called finishing time) of Ji according to the actual schedule of J. Let f + (Ji ) and f (Ji ) be the completion times of Ji
according to the maximal schedule and minimal schedule of J, respectively. We say that Ji is
completion-time predictable if f (Ji ) f (Ji ) f + (Ji ). The execution of Ji is predictable,
or simply Ji is predictable, if Ji is both start-time and completion-time predictable. The execution behavior of the entire set J is predictable if every job in J is predictable. Looking at
Figure 48 again, we see that f (J4 ) is 20, but f + (J4 ) is 16. It is impossible for the inequality
20 f (J4 ) 16 to hold. Therefore, J4 is not completion-time predictable, and the system is
not predictable.
In general, whether a set of jobs has predictable execution behavior depends not only
on the parameters of jobs in the set but also on the algorithm used to schedule the jobs.
For example, while the execution of independent, preemptable but nonmigratable jobs is not
predictable in general, as exemplified by the jobs in Figure 48, it is predictable when the
priorities are assigned on the FIFO basis [Ha]. In Chapters 69, we will use the following
fact, which is true for all priority-driven scheduling algorithms.
T HEOREM 4.4. The execution of every job in a set of independent, preemptable jobs
with fixed release times is predictable when scheduled in a priority-driven manner on
one processor.
Proof. That the highest priority job J1 is predictable is obvious: It always starts at its
release time, and its maximum execution time is larger than its actual execution time.
Suppose that all the i 1 highest priority jobs J1 , J2 , . . . , Ji 1 are predictable. We now
show by contradiction that Ji , which has a lower priority than they but a higher priority
than all the other jobs in the system, is also predictable.
Suppose that s (Ji ) s(Ji ) s + (Ji ) is not true. In particular, we suppose that
s(Ji ) > s + (Ji ). Because the scheduler never schedules a job before its release time,
s + (Ji ) is no earlier than the release ri of Ji . Because the scheduling algorithm is prioritydriven, every job whose release time is at or earlier than s + (Ji ) and whose priorities
are higher than Ji has completed by s + (Ji ) according to the maximal schedule. By
induction hypothesis, we know that every such job has completed by s + (Ji ) according
to the actual schedule as well. s(Ji ) > s + (Ji ) means either that the processor is left
idle in the interval [s + (Ji ), s(Ji )] or a job with a priority lower than Ji is scheduled
76
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 76
in this interval. This contradicts the fact that the algorithm used to schedule the jobs is
priority-driven. Hence, s(Ji ) > s + (Ji ) cannot be true.
A similar argument based on the actual and minimal schedules allows us to conclude that s (Ji ) > s(Ji ) cannot be true. In other words, Ji is start-time predictable.
To show that Ji is also completion-time predictable, we note that since s(Ji )
s + (Ji ), there is more available time in the interval [s(Ji ), f + (Ji )] than in the interval
[s + (Ji ), f + (Ji )]. Moreover, the actual execution time of every job is never larger than
the maximum execution time of the job. If Ji remains incomplete at f + (Ji ) according
to the actual schedule while it can complete by this time according to the maximal
schedule, it must be delayed by one or more jobs with priorities lower than Ji executing
in [s(Ji ), f + (Ji )], or the processor must be left idle for some time in this interval. This
contradicts the fact that the scheduling algorithm is priority-driven. We can, therefore,
conclude that f (Ji ) f + (Ji ). A similar argument based on the actual and minimal
schedules tells us that f (Ji ) is never later than f (Ji ), or that Ji is also completiontime predictable.
Theorem 4.4 tells us that it is relatively easy to validate priority-driven, uniprocessor,
static systems when jobs are independent and preemptable. Because the execution behavior
of all the jobs is predictable, we can confine our attention to the maximum execution times
of all the jobs and ignore the variations in execution times when we want to determine their
maximum possible response times. You will see that this is indeed what we will do in most
parts of Chapters 69. Nonpreemptivity and resource contention invariably introduce unpredictability in execution. Fortunately, it is possible to bound the additional delay suffered by
every job due to these factors reasonably accurately. We will describe the methods for doing
so as well.
4.8.3
Validation Algorithms and Their Performance

The validation problem has many variants, and that of validating static priority-driven systems
is an important variant. Recent advances in real-time scheduling and schedulability analysis
have lead to several sufficient conditions and analytical bounds. They are solutions to this variant and the subjects of discussion in Chapters 69. These theoretical results form a rigorous
basis of good validation algorithms and tools for a wide spectrum of static systems. (A validation algorithm allows us to determine whether all jobs in a system indeed meet their timing
constraints despite scheduling anomalies.) While there are mature validation algorithms and
tools for static systems, good validation algorithms for dynamic, priority-driven systems are
not yet available.
Specifically, we say that a validation algorithm is correct if it never declares that all
timing constraints are met when some constraints may not be. The merits of (correct) validation algorithms are measured in terms of their complexity, robustness, and accuracy. A
validation algorithm is good when it achieves a good balance in performance according to
these conflicting figures of merit.
For example, some existing validation algorithms run in constant time or O(n) time,
where n is the number of tasks in the system. They are well suited for on-line acceptance
tests to determine whether the system should admit a new task. More complex ones run in
pseudopolynomial time but have better performance in the other dimensions.
Liu
January 13, 2000 8:46 a.m.
Section 4.9
chap4
page 77
Off-Line versus On-Line Scheduling
77
Every rigorous validation algorithm is based on a workload model. When applied to

a system, the conclusion of the algorithm is correct if all the assumptions of the model are
valid for the system. A validation algorithm is said to be robust if it remains correct even
when some assumptions of its underlying workload model are not valid. The use of a robust
validation algorithm significantly reduces the need for an accurate characterization of the
applications and the run-time environment and, thus, the efforts in analysis and measurement
of the individual applications for the purpose of validating the workload model. We will see in
later chapters that existing validation algorithms based on the periodic task model are robust
to a great extent. Although the model assumes that jobs in each task are released periodically
and execute for an equal amount of time, such a validation algorithm remains correct in the
presence of release-time jitters, variations in job execution time, and other deviations from
periodic behavior. It is only necessary for us to know the ranges of task parameters (e.g.,
the minimum interrelease time and maximum execution time of jobs), which are much easier
to obtain and validate, either by timing analysis or measurement, than the actual values or
probability distributions of the parameters.
Efficiency and robustness can be achieved easily if we are not concerned with the accuracy of the validation test. A validation algorithm is inaccurate when it is overly pessimistic
and declares tasks unable to meet their timing constraints except when system resources are
unduly underutilized. A scheduler using an inaccurate validation algorithm for an acceptance
test may reject too many new tasks which are in fact acceptable. Because most validation algorithms are based on conditions that are sufficient but not necessary, they are all inaccurate
to some degree, which is the price paid for the sake of robustness. The accuracy of a validation
algorithm depends on whether the actual characteristics of the application systems are accurately captured by the underlying workload model. For example, validation algorithms that
are based on the periodic task model are sufficiently accurate for applications, such as digital
control and constant bit-rate voice and video communications, which are well characterized
by the periodic task model but may have poor accuracy when used to validate applications
that have widely varying processor-time demands and large release-time jitters.
4.9
OFF-LINE VERSUS ON-LINE SCHEDULING

In Section 4.1, we mentioned that a clock-driven scheduler typically makes use of a precomputed schedule of all hard real-time jobs. This schedule is computed off-line before the
system begins to execute, and the computation is based on the knowledge of the release times
and processor-time/resource requirements of all the jobs for all times. When the operation
mode of the system changes, the new schedule specifying when each job in the new mode
executes is also precomputed and stored for use. In this case, we say that scheduling is (done)
off-line, and the precomputed schedules are off-line schedules.
An obvious disadvantage of off-line scheduling is inflexibility. This approach is possible only when the system is deterministic, meaning that the system provides some fixed set(s)
of functions and that the release times and processor-time/resource demands of all its jobs are
known and do not vary or vary only slightly. For a deterministic system, however, off-line
scheduling has several advantages, the deterministic timing behavior of the resultant system
being one of them. Because the computation of the schedules is done off-line, the complexity of the scheduling algorithm(s) used for this purpose is not important. Indeed, as we will
78
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 78
see in the next chapter, complex heuristic algorithms are typically used to find good off-line
schedules that can make nearly full use of the resources.
Competitiveness of On-Line Scheduling. We say that scheduling is done on-line, or
that we use an on-line scheduling algorithm, if the scheduler makes each scheduling decision
without knowledge about the jobs that will be released in the future; the parameters of each
job become known to the on-line scheduler only after the job is released. The priority-driven
algorithms described earlier and in subsequent chapters are on-line algorithms. In Chapter
2 we talked about the admission of each new task depending on the outcome of an acceptance test that is based on the parameters of the new task and tasks admitted earlier. Such an
acceptance test is on-line.
Clearly, on-line scheduling is the only option in a system whose future workload is
unpredictable. An on-line scheduler can accommodate dynamic variations in user demands
and resource availability. The price of the flexibility and adaptability is a reduced ability for
the scheduler to make the best use of system resources. Without prior knowledge about future
jobs, the scheduler cannot make optimal scheduling decisions while a clairvoyant scheduler
that knows about all future jobs can.
As a simple example, suppose that at time 0, a nonpreemptive job J1 with execution
time 1 and deadline 2 is released. An on-line scheduler has two options at time 0: It either
schedules J1 to start execution at time 0 or it postpones the execution of J1 to some later time.
Suppose that the on-line scheduler decides to schedule J1 at time 0. Later at time x < 1, a job
J2 with execution time 1 x and deadline 1 is released. J2 would miss its deadline because
it cannot start execution until time 1. In contrast, a clairvoyant scheduler, which knows J2 at
time 0, would schedule J1 to start execution at time 1 and thus allow both jobs to complete in
time. In the second case, the on-line scheduler decides to postpone the execution of J1 until
some later time x < 1. Now suppose that at time x, J3 is released instead of J2 . The execution
time of J3 is 1, and its deadline is 2. It is impossible for the on-line scheduler to schedule both
J1 and J3 so that they complete in time. Again, a clairvoyant scheduler, knowing the future
release of J3 at time 0, would schedule J1 to start execution at time 0 so it can complete both
J1 and J3 on time.
The system is said to be overloaded when the jobs offered to the scheduler cannot be
feasibly scheduled even by a clairvoyant scheduler. When the system is not overloaded, an
optimal on-line scheduling algorithm is one that always produces a feasible schedule of all
offered jobs. The example above shows that no optimal on-line scheduling algorithm exists
when some jobs are nonpreemptable. On the other hand, if all the jobs are preemptable and
there is only one processor, optimal on-line algorithms exist, and the EDF and LST algorithms
are examples.
During an overload, some jobs must be discarded in order to allow other jobs to complete in time. A reasonable way to measure the performance of a scheduling algorithm during
an overload is by the amount of work the scheduler can feasibly schedule according to the
algorithm: the larger this amount, the better the algorithm. The competitive factor of an algorithm captures this aspect of performance. To define this performance measure, we say that
the value of a job is equal to its execution time if the job completes by its deadline according
to a given schedule and is equal to zero if the job fails to complete in time according to the
schedule. The value of a schedule of a sequence of jobs is equal to the sum of the values of
all the jobs in the sequence according to the schedule. A scheduling algorithm is optimal if
Liu
January 13, 2000 8:46 a.m.
Section 4.9
chap4
page 79
Off-Line versus On-Line Scheduling
79
it always produces a schedule of the maximum possible value for every finite set of jobs. An
on-line algorithm has a competitive factor c if and only if the value of the schedule of any
finite sequence of jobs produced by the algorithm is at least c times the value of the schedule
of the jobs produced by an optimal clairvoyant algorithm.
In terms of this performance measure, EDF and LST algorithms are optimal under the
condition that the jobs are preemptable, there is only one processor, and the processor is not
overloaded. Their competitive factors are equal to 1 under this condition. On other hand, when
the system is overloaded, their competitive factors are 0. To demonstrate, let us consider two
jobs. The first one is released at time 0, and its execution time is 2; the deadline is for some
arbitrarily small positive number . At time , a job whose relative deadline is equal to its
execution time e is released. The value achieved by the EDF or LST algorithm is 0, while the
maximum possible value achievable is e.
As it turns out, the EDF and LST algorithms are not the only algorithms with poor
performance when the system is overloaded. In general, all on-line scheduling algorithms
perform rather poorly. The following theorem due to Baruah, et al. [BKMM] gives us the
performance limitation of on-line scheduling when the system is overloaded.
T HEOREM 4.5. No on-line scheduling algorithm can achieve a competitive factor
greater than 0.25 when the system is overloaded.
*Informal Proof of Theorem 4.5. To gain some insight into why this upper bound
of 0.25 is true, we summarize the proof of Theorem 4.5; you can find the complete
formal proof in [BKMM]. Suppose that there is an adversary of the on-line scheduler.
Over time, the adversary creates two kinds of jobs and offers (i.e., releases) them to the
scheduler: major jobs and jobs associated with major jobs. The relative deadline of every
job is equal to its execution time. (In other words, the job has no slack; the scheduler
should either schedule the job immediately after it is released, or discard it.) We name
the major jobs Ji for i = 0, 1, . . . , max in increasing order of their release times and
denote the execution time and release time of each job Ji by ei and ri , respectively. (max
is some positive integer that we will define shortly.) The adversary creates a sequence
of jobs associated with each major job Ji . The execution times of all the associated jobs
in the sequence are equal to some small number > 0, which is negligible compared
with ei . The first job in the sequence associated with each major job Ji is released at
the same time with Ji . If there is more than one associated job in a sequence, each
subsequent associated job is released at the deadline of the previous associated job in
the sequence. The number of associated jobs in the sequence depends on the action of
the on-line scheduler. Specifically, as long as the on-line scheduler chooses to execute
Ji , the adversary continues to create jobs associated with Ji until the deadline di of Ji .
Whenever the scheduler decides to execute a job associated with Ji , the adversary stops
releasing any more associated jobs.
Let us now consider a busy interval which begins at r0 when the adversary releases
the first major job J0 and the first of the sequence of jobs associated with J0 . Depending
on the action of the on-line scheduler, the adversary may release major job Ji for i > 0
at time ri ; the release time ri of the ith major job and the first of its associated jobs
is equal to ri 1 + ei 1 . In other words, Ji is released at units of time before the
deadline of Ji 1 . It is not possible to schedule both Ji 1 and Ji to complete in time. At
80
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 80
the time ri , the scheduler must choose either to discard Ji 1 and start to execute Ji or
continue to execute Ji 1 and therefore discard Ji .
If the scheduler chooses to discard Ji 1 and begins to execute Ji , the adversary
then releases Ji +1 at ri +1 . As long as the scheduler continues to discard the executing
major job each time a new major job is released in order to execute the new job, the
adversary continues to release the next major job units of time before the deadline
of the executing job. This process continues until the major job Jmax is released for
some positive integer max, and the busy interval ends at the deadline of this job. In this
case, the on-line scheduler discards all the jobs but Jmax , and the total value achieved by
the scheduler in this busy interval is emax . In contrast, the clairvoyant scheduler would
schedule the jobs associatedwith all the major jobs before Jmax and then the major job
Jmax and achieve the value max
k=1 ek .
On the other hand, the scheduler may decide to complete Ji and, upon the completion of Ji , execute a job associated with Ji +1 for i < max. In this case, the adversary
stops releasing any major job after Ji +1 . Moreover, it stops releasing jobs associated
with Ji +1 after the first associated job. (Figure 49 shows this situation for i = 2.) The
value achieved in the busy interval by the on-line scheduler is approximately equal to
ei . However, the clairvoyant scheduler would execute all
jobs associated with jobs J0 ,
J1 , . . . , Ji 1 and then the job Ji and achieve a value of ik=0 ei .
Now suppose that the execution time e0 of the first major job J0 is 1, and for
i > 0, the execution time ei of the ith major job is given by
ei = cei 1
i 1
ek
k=0
If we perform the necessary algebraic manipulation, we will find that competitive factor
of the on-line scheduler is either equal to 1/c, if the scheduler
completes Ji followed
by an associate job, or is equal to the ratio of emax to the sum max
k=0 ek , if the scheduler
discards all the major jobs except Jmax . The former is always greater than or equal to
the latter for every positive integer max if c is equal to 4 or more. For c equal to 4,
r0 r1
major
jobs J0
0
r2
r3
J1
J2
J3
12
associated
jobs
0
12
12 +
FIGURE 49 Example illustrating competitiveness of on-line scheduling.
Liu
January 13, 2000 8:46 a.m.
chap4
Section 4.10
page 81
Summary
81
the competitive factor is 0.25. The execution times of the major jobs are equal to 1, 3,
8, 20, 48, . . . . Figure 49 shows the case where the on-line scheduler executes J2 to
completion. It achieves a value of 8 while a clairvoyant scheduler, knowing that all the
jobs shown in the figure will be released at all time, can achieve a value of 32.
The system used in the proof of Theorem 4.5 is extremely overloaded. It is not surprising that the performance of on-line scheduling algorithms is poor when the system is so
overloaded. Intuitively, we expect that some on-line algorithms should perform well when the
system is only slightly overloaded. To state this more precisely, we let x(t, t
) (t
> t) denote
the load ratio of the interval [t, t
] : It is the ratio of the total execution time of all jobs whose
feasible intervals are contained in the interval [t, t
] to the length t
t of the interval. A system is said to have a loading factor X if the load ratio of the system is equal to or less than X
for all intervals. Our intuition says that if the loading factor of a system is 1 + for some very
small positive number , there should be on-line algorithms whose competitiveness factors
are close to 1. Unfortunately, our intuition fails us. Baruah, et al. showed that the competitiveness factor of an on-line scheduling algorithm is at most equal to 0.385 for any system whose
loading factor is just slightly over 1.
The results on competitiveness of on-line algorithms tell us that when scheduling is
done on-line, it is important to keep the system from being overloaded using some overload
management or load shedding algorithms. Most overload management algorithms take into
account the criticality factors of jobs, not just their timing parameters, when choosing jobs to
be discarded. We will describe a few of them in later chapters.
4.10
SUMMARY
This chapter gave a brief overview of the clock-driven, weighted round-robin and prioritydriven approaches to scheduling. They are the subjects of in-depth discussion of the next
few chapters. This chapter also discussed several important facts about the priority-driven
approach. We need to keep them in mind at all times.
An algorithm for scheduling hard real-time jobs is optimal if it can produce a feasible
schedule as long as feasible schedules of the given jobs exist, that is, when the system is not
overloaded. This is the criterion of merit we use most of the time in this book. The EDF
(Earliest-Deadline-First) algorithm is optimal for scheduling preemptable jobs on one processor. LST (Least-Slack-Time) algorithm is also optimal for preemptable jobs on one processor,
but it requires information on the execution times of all jobs while the EDF algorithm does
not. Neither algorithm is optimal when jobs are nonpreemptable or when there is more than
one processor.
Another important concept is predictability of the timing behavior of jobs. The execution behavior of a system is predictable if the system exhibits no anomalies. We can conclude
that the jobs in a predictable system can always meet their deadlines if the jobs meet their
deadlines according to the maximal schedule of the system, that is, when every job in the
system executes for as long as its maximum execution time. We have shown that when the
jobs are independent and preemptable and are scheduled on one processor, their execution behavior is predictable. This fact allows us to ignore the variations in job execution times during
82
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 82
validation and work with their maximum execution times. Indeed, this is what we will do in
most parts of subsequent chapters.
In general, systems that use priority-driven scheduling have scheduling anomalies. In
subsequent chapters we will discuss efficient validation algorithms that allow us to verify the
timely completions of jobs despite scheduling anomalies. Such an algorithm is correct if it
never concludes that some job completes in time when the job may fail to do so. The merits of
correct validation algorithms are measured by their efficiency, robustness, and accuracy. These
measures tell us how much time a validation algorithm takes to reach its conclusion on the
schedulability of each job or the entire system, whether its conclusion remains correct when
some assumptions of its underlying model are no longer valid, and whether the algorithm is
overly pessimistic.
Finally, the EDF and LST algorithms are not optimal when the system is overloaded so
some jobs must be discarded in order to allow other jobs to complete in time. In fact, these
algorithms perform poorly for overloaded systems: Their competitiveness factors are equal to
zero. Some kind of overload management algorithm should be used with these algorithms.
4.11
EXERCISES
4.1 The feasible interval of each job in the precedence graph in Figure 4P1 is given next to its name.
The execution time of all jobs are equal to 1.
(a) Find the effective release times and deadlines of the jobs in the precedence graph in Figure
4P1.
(b) Find an EDF schedule of the jobs
(c) A job is said to be at level i if the length of the longest path from the job to jobs that have
no successors is i. So, jobs J3 , J6 , and J9 are at level 0, jobs J2 , J5 , and J8 are at level 1, and
so on. Suppose that the priorities of the jobs are assigned based on their levels: the higher
the level, the higher the priority. Find a priority-driven schedule of the jobs in Figure 4P1
according to this priority assignment.
J 1 (0, 10)
J 4 (1, 6)
J 2 (1, 4)
J 5 (3, 9)
J 7 (1, 12)
J 3 (0, 5)
J 6 (2, 10)
J 8 (1, 12)
J 9 (1, 12)
4.2 (a) The execution times of the jobs in the precedence graph in Figure 4P2 are all equal to 1,
and their release times are identical. Give a nonpreemptive optimal schedule that minimizes
the completion time of all jobs on three processors. Describe briefly the algorithm you used
to find the schedule.
Liu
January 13, 2000 8:46 a.m.
chap4
page 83
Exercises
A
B
E
C
G
83
S
T
(b) If the execution times of jobs are arbitrary rational numbers, can your algorithm be modified
so that it will produce optimal preemptive schedules of such jobs? Explain your answer.
4.3 Prove Theorem 4.3 on the optimality of the LST algorithm for scheduling preemptive jobs on one
processor.
4.4 Consider a system that has five periodic tasks, A, B, C, D, and E, and three processors P1 , P2 ,
P3 . The periods of A, B, and C are 2 and their execution times are equal to 1. The periods of D
and E are 8 and their execution times are 6. The phase of every task is 0, that is, the first job of
the task is released at time 0. The relative deadline of every task is equal to its period.
(a) Show that if the tasks are scheduled dynamically on three processors according to the LST
algorithm, some jobs in the system cannot meet their deadlines.
(b) Find a feasible schedule of the five tasks on three processors.
(c) Parts (a) and (b) allow us to conclude that the LST algorithm is not optimal for scheduling
on more than one processor. However, when all the jobs have the same release time or when
they have the same deadline, the LST algorithm is optimal. Prove this statement.
4.5 [Grah] A system contains nine nonpreemptable jobs named Ji , for i = 1, 2, . . . , 9. Their execution times are 3, 2, 2, 2, 4, 4, 4, 4, and 9, respectively, their release times are equal to 0, and
their deadlines are 12. J1 is the immediate predecessor of J9 , and J4 is the immediate predecessor
of J5 , J6 , J7 , and J8 . There is no other precedence constraints. For all the jobs, Ji has a higher
priority than Jk if i < k.
(a) Draw the precedence graph of the jobs.
(b) Can the jobs meet their deadlines if they are scheduled on three processors? Explain your
answer.
(c) Can the jobs meet their deadlines if we make them preemptable and schedule them preemptively. Explain your answer.
(d) Can the jobs meet their deadlines if they are scheduled nonpreemptively on four processors?
Explain your answer.
(e) Suppose that due to an improvement of the three processors, the execution time of every job
is reduced by 1. Can the jobs meet their deadlines? Explain your answer.
4.6 Consider a system that has two processors and uses an on-line preemptive scheduler to schedule
jobs on the processors. At time 0, three independent jobs, J1 , J2 , and J3 , with execution time 1,
1, and 2, respectively, are released. Their deadlines are 2, 2, and 4, respectively. The scheduler
either schedules a portion of J3 before time 1, or it does not schedule J3 before time 1. We now
consider these two cases.
(a) In case (1), the scheduler schedules a portion of J3 before time 1. Now suppose that two more
independent jobs, J4 and J5 , are released at time 2. Their execution times are both equal to 1,
84
Chapter 4
Liu
January 13, 2000 8:46 a.m.
chap4
page 84
and their deadlines are equal to 2. Show that in this case, the scheduler can no longer feasibly
schedule all the jobs, while a clairvoyant scheduler, which foresees the releases of J4 and J5 ,
would not schedule any part of J3 before time 1 and therefore could feasibly schedule all the
jobs.
(b) In case (2), the on-line scheduler does not schedule any portion of J3 before time 1. Show by
constructing an example that there exist future jobs which the on-line scheduler will not be
able to schedule feasibly, while the jobs could feasibly be scheduled if the on-line scheduler
had scheduled J3 before time 1.
4.7 Consider the set of jobs in Figure 4-3. Suppose that the jobs have identical execution time.
(a) What maximum execution time can the jobs have and still can be feasibly scheduled on one
processor? Explain your answer.
(b) Suppose that the release times of J1 and J2 are jittery. The release time of J1 can be as early as
0 and as late as 3, and the release time of J2 can be as late as 1. How can you take into account
this variation when you want to determine whether the jobs can all meet their deadlines?

Chapter 4

Uploaded by

Copyright:

Available Formats

Chapter 4

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chapter 4

Uploaded by

Copyright:

Available Formats

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Commonly Used Approaches

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Weighted Round-Robin Approach

WEIGHTED ROUND-ROBIN APPROACH

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Commonly Used Approaches to Real-Time Scheduling

J1,1 & J2,1

J1,2 & J2,2

Example illustrating round-robin scheduling of precedence-constrained jobs.

Chapter 11 we will describe in detail the application of weighted round-robin algorithms to

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Commonly Used Approaches to Real-Time Scheduling

FIGURE 42 Example of priority-driven scheduling. (a) Preemptive (b) Nonpreemptive.

In general, however, nonpreemptive scheduling is not better than preemptive scheduling.

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Effective Release Times and Deadlines

DYNAMIC VERSUS STATIC SYSTEMS

EFFECTIVE RELEASE TIMES AND DEADLINES

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Commonly Used Approaches to Real-Time Scheduling

Example of effective timing constraints.

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Optimality of the EDF and LST Algorithms

OPTIMALITY OF THE EDF AND LST ALGORITHMS

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Commonly Used Approaches to Real-Time Scheduling

Transformation of a non-EDF schedule into an EDF schedule.

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Optimality of the EDF and LST Algorithms

Example illustrating the LRT algorithm.

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Commonly Used Approaches to Real-Time Scheduling

NONOPTIMALITY OF THE EDF AND THE LST ALGORITHMS

J3 misses its deadline

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Nonoptimality of the EDF and the LST Algorithms

J3 misses its deadline

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.

Commonly Used Approaches to Real-Time Scheduling

CHALLENGES IN VALIDATING TIMING CONSTRAINTS IN PRIORITY-DRIVEN SYSTEMS

Anomalous Behavior of Priority-Driven Systems

Integre Technical Publishing Co., Inc.

January 13, 2000 8:46 a.m.