Scheduling Algorithms for Real-Time Systems

Fredrik Lindh Thomas Otnes Jessica Wennerström

Master Program in Computer Master Program in Computer Master Program in Computer
Engineering Engineering Engineering
Mälardalens University, Mälardalens University, Mälardalens University,
Sweden Sweden Sweden
[email protected] [email protected] [email protected]

ABSTRACT Other overviews of real-time scheduling algorithms have been

Scheduling algorithms are a governing part of real-time sys- presented by Burns[1], Burns and Audsley[2] and by Mo-
tems and there exists many different scheduling algorithms hammadi and Akl[3]. Those are somewhat more in depth
due to the varying needs and requirements of different real- on some topics then this overview.
time systems. The choice of algorithm is important in every
real-time system and is greatly influenced by what kind of The organization of this paper is as follows. In section 2
system the algorithm will serve. Architectures of today cover basic concepts of real-time system and scheduling are ex-
both uniprocessor and multiprocessors which have their own plained. Section 3 presents the scheduling algorithms being
challenges and complexities. In this paper we present an the main section of this paper. It covers static and dynamic
overview of some scheduling algorithms for real-time systems uniprocessor and multiprocessor algorithms presenting some
on both uniprocessor and multiprocessor systems and men- algorithms in each area. The final section, 4, covers sum-
tion some that are optimal in each category. This overview mary and conclusions.
also includes minor explanations of important concepts in
real-time systems and scheduling.
2. BASIC CONCEPTS
1. INTRODUCTION A scheduling algorithm can be seen as a rule set that tells
the scheduler how to manage the real-time system, that is,
The scheduling algorithm is of paramount importance in a
how to queue tasks and give processor-time. The choice of
real-time system to ensure desired and predictable behavior
algorithm will in large part depend on whether the system
of the system. Within computer science real-time systems
base is uniprocessor, multiprocessor or distributed.
are an important while often less known branch. Real-time
systems are used in so many ways today that most of us use
A uniprocessor system can only execute one process at a
them more than PCs, yet we do not know or think about it
time and must switch between processes, for which reason
when we use the devices in which they reside. Cars, planes
context switching will add some time to the overall execution
and entertainment systems are just some devices in which
time when preemption is used.
real-time systems reside, governing the workings of that de-
vice while we do not consider that such a system exists
A multiprocessor system will range from multi-core, essen-
within the chosen device. A real-time-system is a computer
tially several uniprocessors in one processor, to several sep-
system in which the key aspect of the system is to perform
arate uniprocessors controlling the same system.
tasks on time, not finishing too early nor too late. A classic
example is that of the air-bag in a car; it is of great impor-
A distributed system will range from a geographically dis-
tance that the bag inflates neither too soon nor too late in
persed system to several processors on the same board. In
order to be of aid and not be potentially harmful.
a distributed system the nodes are autonomous while in a
multiprocessor system they collaborate somewhat more, but
The choice of algorithm can greatly influence the behavior
this line is not as clear cut as it may sound as similar com-
of a real-time system and for this reason there are many
munication delays will occur.
available algorithms. For the different categories of real-
time systems there are specialized algorithms developed and
In real-time systems processes are referred to as tasks and
in this paper we will attempt to give an overview of many
these have certain temporal qualities and restrictions.[4] All
of the available real-time algorithms.
tasks will have a deadline, an execution time and a release
time. In addition there are other temporal attributes that
may be assigned to a task. The three mentioned are the
basic ones. The release time, or ready time is when the task
is made ready for execution. The deadline is when a given
task must be done executing and the execution time is how
long time it takes to run the given task. In addition most
tasks are recurring and have a period in which it executes.
Such a task is referred to as periodic. The period is the time
from when a task may start until when the next instance of
the same task may start and the length of the period of a Preemption is when a task that is executing on the proces-
task is static. sor becomes interrupted, its state is saved and then it is
exchanged for another task. This switching of tasks on a
An example, shown in Figure 1, of scheduling can be made processor is referred to as a context switch and takes a small
using three tasks T1, T2 and T3 with execution time and amount of time each time it occurs. A preempted task will
deadline of (1, 3), (4, 9) and (2, 9) respectively and periods get to finish executing after the preempting task is done.
equal to their deadlines. These tasks can be scheduled so Not all processors support preemption and algorithms can
that all tasks get to execute before the deadlines. be divided into preemptive and non-preemptive scheduling
algorithms. In real-time scheduling preemption is governed
by priority.

A real-time system also have requirements based on dead-

line, a real-time system can either be hard or soft depending
on the consequences of missing a deadline.

A hard real-time system is never allowed to miss a deadline

Figure 1: Scheduling of T1,T2 and T3.
because that can lead to complete failure of the system. A
hard real-time system can be safety-critical and this means
A system can consist of many tasks and the example above that if a deadline is missed it can lead to catastrophically
uses three periodic tasks, but there can also be aperiodic consequences which can harm persons or the environment.
tasks which are tasks without a set release-time. These tasks It is a crucial requirement that a task starts on time and do
are activated by some event that can occur at more or less not miss the deadline; being not too early nor too late.
any time or maybe even not at all. The scheduling example
shows a minor system and the schedule can be made either In a soft real-time system a deadline is allowed to be missed,
before the system is activated, which is referred to as off-line while there is no complete failure of the system it can lead
scheduling, or during the running of the system and is then to decreased performance.
referred to as online scheduling. The example is very simple
as it does not show priorities or use preemption. There are
also other properties of interest when looking at scheduling.
3. SCHEDULING ALGORITHMS
The scheduling algorithms can be divided into off-line schedul-
Properties a task may use briefly explained:
ing algorithms and online scheduling algorithms. In offline
scheduling all decisions about scheduling is taken before the
system is started and the scheduler has complete knowledge
about all the tasks. During runtime the tasks are executed
in a predetermined order. Offline scheduling is useful if we
have a hard real-time system with complete know-ledge of
all the tasks because then a schedule of the tasks can be
made which ensures that all tasks will meet their deadlines,
if such a schedule exists.

In online scheduling the decisions regarding how to sched-

ule tasks are done during the runtime of the system. The
scheduling decisions are based on the tasks priorities which
Figure 2: The basic temporal attributes of a task. are either assigned dynamically or statically. Static priority
driven algorithms assign fixed priorities to the tasks before
the start of the system. Dynamic priority driven algorithms
assign the priorities to tasks during runtime.
• Release/ready time: The time a task is ready to run
and just waits for the scheduler to activate it.
3.1 Uniprocessor
• Deadline: The time when a task must be finished ex- Scheduling algorithms on uniprocessors have to ensure that
ecuting. all tasks in the system are given enough execution time at
certain points in time to meet their deadlines if possible.
• Execution/run time: The active computation time a Even if the algorithms presented in the following section
tasks need to complete. only schedules processor time, there exists algorithms taking
other resources into account.
• Worst Case Execution Time (WCET): The longest
possible execution time for a task on a particular type
of system.
3.1.1 Static algorithms
Rate monotonic (RM) scheduling algorithm[5, 6] is a unipro-
• Response time: The time it takes a task to finish exe- cessor static-priority preemptive scheme. The algorithm is
cution. Measured from release time to execution com- static-priority in the sense that all priorities are determined
pletes, including preemptions. for all instances of tasks before runtime. What determines
the priority of a task is the length of the period of the re-
• Priority/weight: The importance given a task in con- spective tasks. Tasks with short period times are assigned
text of the schedule at hand. higher priority. RM is used to schedule periodic tasks.
The following are preconditions for the rate monotonic al- This is called priority inversion, when a high priority task in
gorithm formalized by Liu and Layland [5]. essence becomes the task with the lowest priority. Different
protocols for solving this problem are explained next.

1. Periodic tasks have constant known execution times Priority inheritance protocol (PIP)[8] dynamically changes
and are ready for execution at the beginning of each the priority of a task if it is blocking a higher priority task.
period(T ). The priority is then set to the priority of the task it is
blocking. This is referred to as inheriting the priority. The
2. Deadlines(D) for tasks are at the end of each period:
amount of times a high priority task can be blocked is the
(D = T ) number of semaphores it will use. So if a task will use m
semaphores it can in worst case be blocked m times per exe-
3. The tasks are independent, that is, there is no prece- cution. A problem with this protocol is that it does not pre-
dence between tasks and they do not block each other. vent deadlock. Deadlock is when some task holds at least one
semaphore and still tries to lock another semaphore taken
4. Scheduling overhead due to context switches and swap- by another task in a similar state and this will cause the
ping etc. are assumed to be zero. system to halt.

Priority ceiling protocol (PCP)[8] avoids deadlock by having

There exists a scheduling algorithm similar to RM called
each semaphore assigned a priority ceiling which is the pri-
deadline monotonic (DM)[7], used when D < T which al-
ority of the highest priority task that uses it. There are also
lows us to see RM as a specific case of DM. In the case
two restrictions that a task needs to fulfill which is; not hold
of the DM algorithm the deadline determines the priority
semaphores between instances and lock semaphores ”pyra-
of the task; the shorter the deadline the higher the prior-
midically”. To be able to preempt a task must have higher
ity. Both RM[5] and DM[7] are optimal static scheduling
priority than all used semaphores. Any holding task will
algorithms. It has been shown[5] that for an optimal static
inherit the priority of a task becoming blocked but the pri-
priority driven algorithm an upper bound for the processor
ority ceiling will prevent a task from being blocked more
utilization is 70 percent for large task sets.
than once per execution.
Example of RM, see Figure 3: Tasks T1 & T2 with execution
times & periods (1, 4) & (2, 6). T1 with a shorter period & 1. Task T wants to lock semaphore S.
therefore higher priority runs before T2 . They then run as
they are released. 2. If the priority of T is higher than the highest priority
ceiling of all the currently locked semaphores, or if no
semaphores are locked, then T is allowed to lock S ;
else it becomes blocked.

3. If T however is blocked by another semaphore S’ locked

by task T’, T’ then inherits the priority of T.

4. When T’ releases semaphore S’, T gets to execute and

lock S.

Immediate inheritance protocol (IIP) is based on PCP and

Figure 3: Scheduling example of Rate-Monotonic. uses priority ceilings for semaphores but differs in that as
soon as a semaphore is taken by a task that task inherits
the priority of the semaphore. In both PCP and IIP, in
3.1.2 Extensions of RM contrast to PIP, deadlock is prevented and a task can only
There exist some extensions for RM to make it a more useful be blocked one time. This means that if a task will use m
algorithm. The extensions make it possible to use RM when semaphores it can in worst case be blocked one time per
tasks share resources. To prevent simultaneous use of shared execution.
resources a technique called semaphore, which is a sort of
lock, is used which a task must take before entering a critical 3.1.3 Dynamic algorithms
section and then must release after leaving the critical sec- Earliest deadline first (EDF)[2, 5] is a dynamic priority
tion. A critical section is a piece of code accessing a shared driven scheduling algorithm which gives tasks priority based
resource. When using semaphores problems with blocking on deadline. The preconditions for RM are also valid for
can occur. Blocking is when a task is prevented from exe- EDF, except the condition that deadline need to be equal to
cuting by tasks with lower priority. This occurs when a task period. The task with the currently earliest deadline during
has already gained access to a semaphore and a task with runtime is assigned the highest priority. That is if a task is
higher priority wants it. Then the task with higher priority executing with the highest priority and another task with an
needs to wait until the semaphore becomes available. Dur- earlier deadline becomes ready it receives the highest prior-
ing that time other tasks can preempt the lower priority task ity and therefore preempts the currently running task and
and execute as long as they do not require that semaphore begins to execute. EDF is an optimal[5, 9] dynamic priority
and thus the higher priority task is delayed even further. driven scheduling algorithm with preemption for a real-time
system on a uniprocessor. EDF is capable of achieving full Assigning the server the highest priority is preferable if we
processor utilization[5]. want good responsiveness and ensure that aperiodic tasks
meet their deadlines, while assigning the DS medium prior-
Least laxity first (LLF)[2, 10], also known as least slack time, ity aperiodic tasks might miss their deadlines as other tasks
is a dynamic priority driven scheduling algorithm that as- might preempt the server. A drawback is that the algorithm
signs priority based on the laxity. The definition of laxity is violates the first assumption of RM since it allows the server
the tasks deadline minus the remaining computation time of to defer its execution, thus it may not be ready to execute
the task. It can also be described as the maximum time a at the beginning of the period.
task can wait before it needs to execute in order to meet its
deadline. The task with the currently least laxity is assigned An example of DS illustrated by Figure 4. The tasks, T1 &
the highest priority and is therefore executed. The executing T2 , have the attributes presented in the RM example. The
task will be preempted by any other task with a currently deferrable server is assigned medium priority. At time two
smaller laxity. When the task is executing the laxity re- an aperiodic request arrives. Since the server has medium
mains constant. If two tasks have similar laxity they will priority, T2 is preempted and the server serves the request.
continually preempt each other and it will therefore create The server capacity is exhausted by the aperiodic task which
many context switches. But if we ignore the cost of the con- executes for two time units. T1 becomes ready at time four
text switches LLF is also, as EDF, an optimal[10] dynamic and executes for one time unit delaying the execution of
scheduling algorithm. T2 even further. At time five T2 is allowed to finish and
the server replenishes its capacity. In the second period of
T2 there are no higher priority tasks ready so it is allowed
3.1.4 Serving aperiodic tasks to execute immediately after release. After T2 has finished
executing, an aperiodic request occurs. The server is not
There are systems where algorithms for scheduling of both
allowed to serve this request since T1 is also ready. When
periodic and aperiodic tasks are needed. Due to the unpre-
T1 is finished the server will serve the request.
dictable arrival of aperiodic tasks it is harder to guarantee
a response time. The server algorithms presented improve
the average response time for aperiodic tasks and are based
on RM, although similar approaches based on EDF exists.
One type of task that is generally not handled well is spo-
radic tasks. A sporadic task is an aperiodic task with a hard
deadline, WCET and a known minimum time between per-
mitted arrivals of the task called interarrival time. Assigning
a server high priority will improve its response time.

The background server algorithm[11] is one simple approach

for servicing soft deadline aperiodic tasks. If there are ready
periodic tasks they get priority over aperiodic tasks to run;
only when the processor is idle, no periodic tasks are execut-
ing, aperiodic arrivals are served in the background. This
implies that a high load of periodic tasks result in bad re-
sponse times for aperiodic tasks.

The polling server (PS) algorithm[11] is another aperiodic

Figure 4: Scheduling example of Deferrable Server.
server algorithm which is a bit more sophisticated. Aperi-
Ci is execution time and Ti is the period.
odic tasks are handled by a periodic task, the PS. Arrivals
of aperiodic tasks are served at the beginning of each of
the PS’s periods, if there are no request pending the server
suspends itself allowing periodic tasks to execute. A late ar- The priority Exchange server (PE)[11] is another capacity
rival, after suspension of the PS, will be queued and served preserving scheduling algorithm that uses a periodic server.
at the next task invocation. The algorithm resembles the DS algorithm but differs in how
capacity is preserved. The server replenishes its capacity at
The deferrable server (DS) algorithm[6] is an algorithm with the beginning of each period. Aperiodic tasks waiting to
focus on quick response times for aperiodic tasks. The algo- be served at the start of a new period will be executed at
rithm improves the average response time for aperiodic tasks the priority of the server and consume capacity, assuming
compared to the background and polling server. The DS is a the server currently has the highest priority. On the other
special kind of periodic task with period and capacity, which hand, if there are no aperiodic tasks ready for execution the
serves aperiodic tasks and can do so until available capac- server allows a ready lower-priority periodic task to execute
ity runs out, or the end of the period is reached. Regular in exchange for accumulation of capacity at the priority level
periodic tasks are ready at set times whereas the DS may of that periodic task. Whenever an aperiodic task requests
receive requests at anytime during the period and therefore to run capacity available for the server at the task with the
executes at different times. When the DS serves a request highest priority, amongst those with which capacity has been
capacity is consumed else capacity is preserved. However, exchanged, will be consumed by the aperiodic task. This
any remaining capacity at the end of the period is lost and type of capacity exchange continues on lower levels with pe-
the DS is replenished at the beginning of the next period. riodic tasks, server capacity will therefore not be lost, just
stored at lower priority levels or consumed by aperiodic re- 3.2 Multiprocessor
quests unless at some point the capacity is exchanged with In multiprocessor scheduling, algorithms developed for unipro-
the idle task. cessor scheduling can be applied if we consider each core of
the multiprocessor as an isolated core, which is a unipro-
The DS algorithm is not as complex as the PE algorithm due cessor. However, in multiprocessor scheduling the difficulty
to the way capacity is preserved at the priority of the server. of verifying that the execution of different tasks on multiple
The price to pay for this simplicity is a lower schedulability cores does not interfere with each other and also determin-
bound than PE. In order for both algorithms to function ing which tasks should be given to a certain core increases
properly particular resource utilization has to be reserved the complexity greatly compared to uniprocessor schedul-
for the server. The server utilization, US , which is the ratio ing. Moreover scheduling algorithms for multiprocessors of-
of execution time to the period, directly affects the schedu- ten involves heuristics to simplify the task of finding a fea-
lability of the system. The highest utilization bound for sible schedule. The two main approaches for scheduling al-
periodic tasks at which the periodic tasks can be scheduled, gorithms on multiprocessors are global scheduling and par-
UP , is determined by RM. titioning scheduling.
US + 2
DS : UP = ln Global scheduling algorithms put ready tasks in a queue
2US + 1
sorted based on priority. The task with currently highest
2 priority, which is first in the queue, is selected by the sched-
P E : UP = ln uler and will execute on one of the processors and may be
US + 1
preempted or migrated if necessary.
From the equations we see that for any given value, Us,
where 0 < US < 1, the schedulability bound, UP , is lower In partitioning scheduling algorithms each task is assigned
for the DS algorithm than in the case of the PE algorithm. to one processor and will exclusively execute on that proces-
Another implication is that for a given Up the server utiliza- sor. This will result in that instead of being a multiprocessor
tion is lower for the DS algorithm than for the PE algorithm. problem it will be a set of uniprocessor problems. Unipro-
cessor algorithms can then be used for each processor which
From a server perspective scheduling sporadic tasks means is an advantage, but the problem of partitioning the task
that the server is required to have a period less or equal to to the processors is hard and usually solved by using non-
the interarrival time of the sporadic task and not consume optimal heuristic. Also as shown by [13] there exist systems
more execution time than the capacity available. which are schedulable if and only if tasks are not partitioned.
The sporadic server (SS)[11], like the DS, consists of a pe- The myopic algorithm[12] is a multiprocessor algorithm that
riodic server for aperiodic tasks but how it replenishes ca- schedules tasks not only for CPU time but also for shared
pacity differs. The server determines at which time in the resource requirements and uses heuristics to simplify the
future capacity will be replenished depending on when aperi- search for a feasible schedule at any given time. As a new
odic requests occur and the priority of the current executing task arrives it will be given to one of the processors which
periodic tasks. will see if schedulability of that task can be guaranteed. If
no schedule can be found the task will be sent on to another
The explanation of the algorithm involves the following terms: processor, while it will be kept if a schedule is found that can
guarantee that task along with all currently existing tasks
• Pexe the priority level of the currently executing task on that processor. When a new task is made ready this
in the system. algorithm looks at:

• PS the priority level of the sporadic server.

• Arrival time TA
• Active is used to describe the priority level when Pexe
> PS . • Deadline TD

• Idle is the opposite of active, Pexe < PS . • Worst case processing time TP
• RTS is the time at which the server replenishes con- • Resource requirements TR
sumed capacity.
Premises for tasks are that they are independent, nonperi-
The sporadic server starts with fully replenished capacity. odic, non-preemptive and use resources in shared or exclu-
Whenever the server becomes active RTS is set to the cur- sive mode. When scheduling a task the algorithm calculates
rent time added to the period of the server. How much the earliest start time, TEST for the task based upon when
the server should replenish at RTS is determined when the the required resources will be available. The following con-
server becomes idle or all the capacity has been consumed. dition is to be true for every task when a new schedule is
The amount to replenish is the capacity consumed from the made:
point the server was activated to the point it becomes idle
0 < TA < TEST < (TD − TP )
or runs out of capacity. The sporadic server performs bet-
ter than the background server and the polling server. The All tasks that are at any one point to be scheduled is placed
performance is comparable to the DS and PE algorithms al- in a list sorted according to deadline and a schedule is searched
though they are in some cases inferior to the SS algorithm. for by using a tree structure to find feasible schedules in
which the root is an empty schedule and each leaf is a sched- server algorithms have a drawback in that they violate a
ule while not every leaf is feasible. The tasks are inserted requirement of the rate monotonic algorithm stating that a
into the schedule, which then is one node level down in the task should be ready to execute at the beginning of every
tree, one at a time until a schedule is found or a deadline period. Yet, servers offer an advantage when dealing with
miss will occur. In the case of a miss there will be some back aperiodic tasks. Another kind of extension algorithms are
tracking in the tree and a new schedule will be looked at and protocols of which some are mentioned in this paper. They
so forth until a plausible schedule is found. Every time an- aid in dealing with mutual exclusion and deadlocks.
other task is to be inserted there is a heuristic function H,
which looks at more than deadline, is called to evaluate a This shows that the complexities on uniprocessor systems
portion of the tasks in the list to find the most appropriate can be quite high and yet it is far greater on multiproces-
task to insert into the schedule at that point. sor systems which are becoming the commonplace even in
small real-time systems. As complexity increases utilization
The original version of this algorithm looked at all tasks that decreases and on multiprocessor systems with static priority
were not in the schedule every time a task was to be added, scheduling it can be at most 50%[17].
but this algorithm only looks to the first few in a sorted
list and this near sightedness is what makes the algorithm Multiprocessor systems are the future as we see it now, but
myopic. Compared to the non-myopic version of this algo- finding algorithms that takes full advantage of these systems
rithm and to other multiprocessor algorithms of its time the is an arduous task in which much effort has been and is
myopic algorithm was an efficient improvement. The my- being made by researchers. Future work could be to focus
opic algorithm is O(nk) where 1 ≤ k ≤ n depending on the on these new algorithms being produced as well as dynamic
number of tasks to be scheduled at the time as n equals all based server algorithms.
tasks to be scheduled and k is a subset there of. This then
makes the myopic algorithm have O(n) while the original
non-myopic algorithm was O(n2 ).
5. REFERENCES
[1] Burns A.,”Scheduling hard real-time systems: a review”
Software Engineering Journal, May 1991.
Pfair scheduling algorithms[14] are based upon the idea of
giving each task access to resources in proportion to the [2] Burns A. and Audsley N., ”REAL-TIME SYSTEM
demand for resources it has in comparison to the other tasks SCHEDULING” Predicatably Dependable Computer
scheduled. Tasks are given a new attribute called rational Systems, Volume 2, Chapter 2, Part II. or Department of
weight (x.w) which is defined as execution requirement (x.e) Computer Science, University of York, UK.
divided by period (x.p), in which x.e and x.p are integer [3] Mohammadi A. and Akl S. G., ”Scheduling Algorithms
values and x.p > 1. To be schedulable 0 < x.w < 1 must be for Real-Time Systems”, Technical Report No. 2005-499,
true for any task and the sum of all tasks x.w < m, where School of Computing, Queen’s University Kingston,
m is the sum of all resources. Tasks are then split into Ontario Canada K7L 3N6, July 15, 2005.
subtasks, that is points where a task may be preempted so [4] Cottet F., Delacroix J. and Mammeri Z., ”Scheduling in
another will get access to resources for awhile, and a function Real-Time Systems”, published by John Wiley & Sons
is used to determine how to schedule these subtasks such Ltd, Chichester, West Sussex, England, ISBN
that all requirements are met. In addition there are several 0-470-84766-2, year 2002.
rules governing how this is done correctly to get the greatest [5] Liu C.L. and Layland J.W., ”Scheduling Algorithms for
benefit from this approach. There are at least three known Multiprogramming in a Hard-Real-Time Environment”
optimal pfair algorithms: PF[14], PD[15] and ER-PD[16]. Journal of the Association for Computing Machinery,
vol. 20, no. 1, pp. 46-61., year 1973.
4. SUMMARY AND CONCLUSIONS [6] Strosnider J. K., Lehoczky J. P. and Sha L., ”The
Deferrable Server Algorithm for Enhanced Aperiodic
In the simple case, scheduling may seem straight forward
Responsiveness in Hard Real-Time Environments”, IEEE
and is easy enough to understand, but once more tasks are
Transactions on Computers, vol. 44, no. 1, January 1995.
added it becomes more troublesome to complete a schedule.
Introducing several resources makes for more complicated [7] Leung J. Y.-T., Whitehead J., ”On the complexity of
schedules and increases complexity in the algorithms used fixed priority scheduling of periodic, real-time tasks”,
to create these schedules. The requirements that real-time Performance Evaluation, vol. 2, issue 4, pages 237-250,
systems must fulfill are several and equally many are the December 1982.
approaches how to schedule such systems. As mentioned in [8] Sha L., Rajkumar R. and Lehoczky J. P., ”Priority
this paper optimal algorithms exist but their optimality is Inheritance Protocols: An Approach to Real-Time
often only theoretical and not practical in actual systems. Synchronisation”, IEEE Transactions on Computers
39(9), pp. 1175-1185, September 1990.
The scheduling of aperiodic and sporadic tasks is more dif- [9] Dertouzos M., ”Control robotics: The procedural control
ficult than periodic tasks due to their unpredictability. To of physical processes”, Proc. IFIP Cong., pp. 807-813,
overcome this, algorithms based on server tasks for handling year 1974.
aperiodic arrivals are one solution. Algorithms of this kind [10] Dertouzos M.L. and Mok A.K.L., ”Multiprocessor
are an extension to the scheduling algorithm governing the On-Line Scheduling of Hard Real-Time Tasks” IEEE
system. All servers presented in this paper are based upon Transactions on Software Engineering, vol. 15, no. 12,
systems working on the rate monotonic algorithm, but there December 1989.
are other server algorithms extending systems based on the [11] Sprunt B., ”Aperiodic Task Scheduling for Real-Time
earliest deadline first algorithm. Several of the presented Systems” Ph.D. Dissertation, Department of Electrical
 and Computer Engineering, Carnegie Mellon University,
August 1990.
[12] Ramamritham K., Stankovic J. A. and Shiah P.-F.,
”Efficient Scheduling Algorithms for Real-Time
Multiprocessor Systems”, IEEE Transactions on Parallel
and Distributed Systems, vol. 1, no. 2, April 1990.
[13] Carpenter J., Funk S., Holman P., Srinivasan A.,
Anderson J. and Baruah S., ”A Categorization of
Real-time Multiprocessor Scheduling Problems and
Algorithms”, Handbook of Scheduling: Algorithms,
Models, and Performance Analysis, Edited by J. Y.
Leung, Published by CRC Press, Boca Raton, FL, USA,
year 2004.
[14] Baruah S.K., Cohen N., Plaxton C.G., and Varvel D.,
”Proportionate progress: A notion of fairness in resource
allocation.”, Algorithmica, 15:600-625, year 1996.
[15] Baruah S.K., Gehrke J. and Plaxton C.G., ”Fast
scheduling of periodic tasks on multiple resources”,
Proceedings of the 9th International Parallel Processing
Symposium, p. 280-288, April 1995.
[16] Anderson J. H. and Srinivasan A., ”Early-Release Fair
Scheduling”, Department of Computer Science,
University of N. Carolina, Chapel Hill, NC 27599-3175,
year 2000.
[17] Andersson B. and Jonsson J., ”The utilization bounds of
portioned and pfair static-priority scheduling on
multiprocessors are 50%”, Department of Computer
Engineering, Chalmers University of Technology, SE-412
96 Göteborg, Sweden, year 2003.