Scheduling Algorithm

Concept of Multiprogramming
 Multiprogramming takes
advantages of the fact that
processes will spend a great deal of
their time waiting for I/O operations
to finish.
 While process #1 is waiting for I/O,
the CPU will execute process #2.
CPU-I/0 Burst Cycle
 A program running on a
computer has main different
“bursts” of activities: bursts
of CPU activity and bursts of
Input/Output activity.
 Since this is cyclic in nature,
it is called the CPU-I/O
Burst Cycle.
 A process is a series of CPU and I/O Bursts.

Load
Store
Add
Store
Read from File Load
Store
Add
Write to File

Store
Increment Index
Write to File
Process Spends most of its time
waiting for I/O
Three Process Executing without
Multiprogramming
Three Process Executing with
Multiprogramming
 Scheduling Criteria
 Performance criteria – What do we
want from our scheduling algorithm?
 Utilization – As near 100% CPU time
as possible.
 Throughput – Highest possible
number of finished processes per unit
time.
 Turnaround time – As low as
possible for jobs from start to finish.
 Response time – For interactive
systems, this is more important than
turnaround time.
 Waiting Time – minimize time in
ready queue and device queues.
Preemptive vs. Nonpreemptive
Scheduling

 Preemptive scheduling – processes

using the CPU can be removed by the
system.
 Nonpreemptive scheduling –
processes using the CPU cannot be
removed by the CPU.
 Starvation – A situation that arises
when a process never gets to the CPU
(or to perform an I/O operation, etc.)
 Scheduling Algorithms

 Include:
First-Come-First-Served
Shortest Job First
Round Robin
Priority
Guaranteed
Lottery
Real-Time
First-Come-First-Served
 Algorithm is the simplest CPU
scheduling.
 Whichever process requests the CPU
first gets first.
 It is implemented using a standard
FIFO single queue.
 Waiting time can be long and it
depends heavily on the order in
which processes request CPU time.
An Example of First-Come-First-Served
An Example of First-Come-
First-Served

Process Burst time (in msec)

 Scenario #
#11for
forFCFS
FCFSScheduling
Scheduling

Processing time =
Scenario #2 for FCFS Scheduling

Waiting time =
Scenario #3 for FCFS Scheduling

Waiting time =
Shortest Job First
 Most appropriately called Shortest
Next CPU Burst First because it
bases the order upon an
approximation of how long what the
next CPU burst will be.
 This can be proven to be the optimal
scheduling algorithm with the shortest
average processing (and waiting) time.
 The SJF algorithm can be preemptive
or non-preemptive, with the
preemptive SJF algorithm more
properly being called shortest-
remaining-time-first scheduling.
 Gantt Chart for Shortest
Job First example

Waiting time =
CPU Burst Length
 The real difficulty is that we are trying
to predict how long the next CPU burst
will be and this cannot be done with
any real accuracy for short-term CPU
scheduling.
 The next burst is usually predicted
using the exponential average of
previously CPU bursts:
Predicting the next CPU burst length

 More recent terms all have more weight than earlier term in the
calculation.
Priority Scheduling
 Involves a priority assigned to each
process, which is scheduled in
accordance with its priority.
 Processes with equal priority are
scheduled on a FCFS basis.
 A SJF algorithm is special case of a
priority scheduling algorithm with
priority(p) being proportional to l/P
Priority Levels
 There is no general agreement on whether
0 is the highest or lowest priority (priority
numbers are assumed to be positive)
- UNIX uses 0 as the highest priority
- IBM’s MVS uses it as the default
(lowest) priority.
Setting Priorities
Priorities can be set:
 Internally – (by some
measurable quantity or
quantities such as time limits,
memory requirements,
number of open files, I/O
burst-to-CPU burst ratio, etc.)
 Externally – (by system
policy, such as process
importance, type or availability
of funds, sponsoring
department, etc.)
Starvation
 Is a major problem of priority
scheduling algorithms.
 On a busy system, a low-priority
process may sit for extremely long
periods of time.
 A solution to the problem is aging,
where we increment the priority
(make it a higher priority) for every
1-15 minutes of waiting.
Scenario For Priority Scheduling
 Gantt Chart for Priority Scheduling Scenario

Processing time =
Round Robin Scheduling
 is designed for time-sharing
system.
 It is similar to the FCFS
scheduling, but preemptive is
added to switch between
processes.
 A time quantum is typically 10
to 100 milliseconds.
 The ready queue is implemented
in FIFO manner.
Round Robin Scheduling and
Preemption
 If a process needs less than a
time quantum, it releases the CPU
voluntarily.
 If a process needs more than a
time quantum, it is preempted
from the CPU and placed at the
back of the ready queue.
Scenario for Round Robin
Scheduling

Process Burst Time

 Gantt Chart for Round Robin
Scheduling Scenario

Processing time =
 Time Quanta
 The performance of round robin
scheduling is extremely dependent
upon the size of the time quantum in
use.
 If the time quantum is large
(approaching infinity), it approaches a
FCFS algorithm.
 If the time is quantum is small, it
appears (in theory at least) that each
user has his/her own processor.
 Time Quanta and Context Switches
 We need the time quantum to be large with
respect to the context-switch time (time
it takes to switch processes) because each of
these can effectively slow the processor.
 Guaranteed Scheduling
 If fairness is an important concern, and there are
n users on a uniprocessor system, each user
should be able to get l/n of the system’s time.
 To make good on this promise to provide each
user with l/n of the CPU time, we keep track of
how much CPU each user has gotten over a time
frame and calculate the ratio of actual CPU time
to entitled CPU time.
 A ratio of 0.5 means that a process should have
gotten half as much CPU time; a ratio of 2.0
means that the process has gotten twice as
much as it should have gotten.
 Such a scheduling algorithm runs the process
with the poorest ratio until it catches up to and
passes its nearest competitor
 Lottery Scheduling
 Every process is given, in effect,
tickets for a lottery, where the prize is
the next time slice (or some other
system resource).
 Applied to CPU scheduling, there may
be 50 lotter drawings each second,
with each winner getting 20msec of
CPU time.
 Important processes can get
extra CPU time by being given extra
“tickets” for the drawings.
 Cooperating processes can exchange
tickets if they wish.
Multilevel Queue Scheduling
Multilevel Feedback Queues:
Shifting to Lower Priorities
Multilevel Feedback Queues:
Shifting to Higher Priorities
Multiple Processor Scheduling
 Process scheduling on a
multiprocessor system is more
complex.
 It is easier to schedule
homogeneous multiprocessor
systems than heterogeneous
systems.
 Identical processors can do load
sharing with separate ready
queues or a common ready
queue.
 Symmetric vs. Asymmetric
Multiprocessing
 In symmetric multiprocessing,
all the processors are considered
peers and any one of them can
handle any sort of task.
 In asymmetric multiprocessing,
there is a hierarchy among the
processors and one of them may
handle the task scheduling
processes for the others.
 What is Real-time Scheduling?
 A real-time system is one in which
time plays a crucial role.
 An example is a CD player which
must read and then translate the
bits into music within a tight time
frame.
 Real-time systems can be hard real
time (where absolute deadlines
must always be met) or soft real
time (where an occasional deadline
can be missed).
 Real-time Scheduling
 Real-time behavior is achieved by dividing the program into a
number of processes, each of which have known behavior.
 Real-time systems react to events which can be periodic
(happening at regular intervals) or aperiodic (not happening
at regular intervals).
 If there are m periodic events and events i occurs with a
period Pi and requires Ci seconds of CPU time, then the load
can only be handled if

Such a system is schedulable.

 Algorithm Evaluation
 There are several ways in which we
can evaluate the scheduling
algorithms:
 Deterministic Modeling
 Queueing Models
 Simulation
 Our goal is to see if they help us
meet the performance criteria that
were discussed earlier.
 Deterministic Modeling
 We will assume a predetermined
set of data.
 Given that data, we will
determine how the scheduling
algorithm will perform.
 Deterministic Modeling is easy
to understand and implement but
it only tells us about the data sets
that we use.
Simulations
 We use random numbers to give
us a large set of data that should
be representative of real-life
processing scenarios.
 The distributions can be defined
either empirically or
mathematically (e.g., a Poisson
distribution).
 Such simulations can be
expensive and more informative.
THANK YOU!