Chapter 5: CPU Scheduling

1
Chapter 5 CPU Scheduling

• Always want to have CPU working
• Usually many processes in ready queue
• Ready to run on CPU
• Consider a single CPU here
• Need strategies for
• Allocating CPU time
• Selecting next process to run
• Deciding what happens after a process completes a
  system call, or completes I/O
• Short-term scheduling
• Must not take much CPU time

2
CPU Burst I/O Burst Alternation (Fig. 5.1)
What happens when the I/O burst occurs? What is
the process state at that time?
3
When does/can scheduling occur?

• When a process switches from running to waiting
  state
• E.g., when it does an I/O request
• When a process terminates
• When a process switches from running to ready
  state
• E.g., when a timer interrupt occurs
• When a process switches from waiting state to
  ready state
• E.g., completion of I/O

Cases 1 2 Non-preemptive scheduling Cases 3
4 Preemptive scheduling (will also include case
1 2)
4
Scheduling Performance Criteria

• CPU utilization
• Percentage of time that CPU is busy (and not
  idle), over some period of time
• Throughput
• Number of jobs completed per unit time
• Turnaround time
• Time interval from submission of a process until
  completion of the process
• Waiting time
• Sum of the time periods spent in the ready queue
• Response time
• Time from submission until first output/input
• Approximate by time from submission until first
  access to CPU

5
Scheduling Algorithms

• First-Come, First-Served (FCFS)
• Complete the jobs in order of arrival
• Shortest Job First (SJF)
• Complete the job with shortest next CPU burst
• Priority (PRI)
• Processes have a priority
• Allocate CPU to process with highest priority
• Round-Robin (RR)
• Each process gets a small unit of time on CPU
  (time quantum or time slice)

6
FCFS First-Come First-Served

• Ready queue data structure a FIFO queue
• Example
• Draw Gantt chart
• Compute the average waiting time for processes
  with the following next CPU burst times and ready
  queue order
• P1 20
• P2 12
• P3 8
• P4 16
• P5 4
• Waiting time Sum of the time periods spent in
  the ready queue

7
Solution Gantt Chart Method
FCFS First-Come First-Served

• Waiting times
• P1 0
• P2 20
• P3 32
• P4 40
• P5 56
• Average wait time 148/5 29.6

8
FIFO

• Disadvantage long waiting times

9
SJF Shortest Job First

• The job with the shortest next CPU burst time is
  selected
• Example (from before)
• CPU job burst times and ready queue order
• P1 20
• P2 12
• P3 8
• P4 16
• P5 4
• Draw Gantt chart and compute the average waiting
  time given SJF CPU scheduling

10
SJF Solution

• Waiting times (how long did process wait before
  being scheduled on the CPU?)
• P1 40
• P2 12
• P3 4
• P4 24
• P5 0
• Average wait time 16

(Recall FIFO scheduling had average wait time of
29.6)
11
SJF

• Provably shortest average wait time
• BUT What do we need to actually implement this?

12
SJF

• Requires future knowledge
• But, may estimate or predict
• How to estimate or predict the future?
• Use history to predict duration of next CPU burst
• E.g., base on duration of last CPU burst and a
  number summarizing duration of prior CPU bursts
• tn1 a tn (1 - a) tn
• Where
• tn is the actual duration of nth CPU burst value
  for the process,
• a is a constant indicating how much to base
  estimate on last CPU burst, and
• tn is the estimate of CPU burst duration for time
  n

13
Example Estimate

• Say, a 0.5
• t0 10 (last estimate)
• Current (measured) CPU burst, t 6
• What is estimate of next CPU burst?
• t1 0.5 6 0.5 10 8

14
Time Complexity of SJF with estimation method?

• Need to keep ready queue ordered
• Process with the shortest estimated next CPU
  burst must be kept at the beginning of the
  queue
• If queue is already sorted, time is O(n) or
  O(lg n) where n is queue length

15
Priority Scheduling

• Have to decide on a numbering scheme
• 0 can be highest or lowest
• Priorities can be
• Internal
• Set according to O/S factors (e.g., memory
  requirements)
• External
• Set as a user policy e.g., User importance
• Static
• Fixed for the duration of the process
• Dynamic
• Changing during processing
• E.g., as a function of amount of CPU usage, or
  length of time waiting (a solution to indefinite
  blocking or starvation)

16
Starvation Problem

• Priority scheduling algorithms suffer from
  starvation (or indefinite blocking)
• In a heavily loaded system, a steady stream of
  higher-priority processes can result in a low
  priority process never receiving CPU time
• I.e., it can starve for CPU time
• One solution aging
• Gradually increasing the priority of a process
  that waits for a long time

17
Which Scheduling Algorithms Can be Preemptive?

• FCFS (First-come, First-Served)
• Non-preemptive
• SJF (Shortest Job First)
• Can be either
• Choice when a new job arrives
• Can preempt or not
• Priority
• Can be either
• Choice when a processes priority changes or when
  a higher priority process arrives

18
RR (Round Robin) Scheduling

• Method
• Give each process a unit of time (time slice,
  quantum) of execution on CPU
• Then move to next process in ready queue
• Continue until all processes completed
• Necessarily preemptive
• Requires use of timer interrupt
• Example
• CPU job burst times order in queue
• P1 20
• P2 12
• P3 8
• P4 16
• P5 4
• Draw Gantt chart, and compute average wait time
• Time quantum of 4
• Assume 0 context switch time

19
Solution

• Waiting times
• P1 60 - 20 40
• P2 44 - 12 32
• P3 32 - 8 24
• P4 56 - 16 40
• P5 20 - 4 16
• Average wait time 30.4

20
Other Performance Criteria

• Response time
• Estimate by time from job submission to time to
  first CPU dispatch
• Assume all jobs submitted at same time, in order
  given
• Turnaround time
• Time interval from submission of a process until
  completion of the process
• Assume all jobs submitted at same time

21
Response Time Calculations
22
Turnaround Time Calculations
23
Performance Characteristics of Scheduling
Algorithms

• Different algorithms will have different
  performance characteristics
• RR (Round Robin)
• Average waiting time often high
• Good average response time
• Important for interactive or timesharing systems
• SJF
• Best average waiting time
• Some overhead w.r.t. estimates of CPU burst
  length ordering ready queue

24
Context Switching Issues

• This analysis has not taken context switch
  duration into account
• In general, the context switch will take time
• Just like the CPU burst of a process takes time
• Response time, wait time etc. will be affected by
  context switch time
• RR (Round Robin) quantum duration
• To lower response times, want smaller time
  quantum
• But, smaller time quantum increases system
  overhead
• To increase throughput, want a large quantum
  (compared to context switch time)

25
Example

• Calculate average wait time for RR (round robin)
  scheduling, for
• Processes
• P1 24
• P2 4
• P3 4
• Assume above order in ready queue P1 at head of
  ready queue
• Quantum 4 context switch time 1

26
Solution Average Wait Time With Context Switch
Time

• P1 0 11 4 15
• P2 5
• P310
• Average 10

(This is also a reason to dynamically vary the
time quantum. E.g., Mach O/S.)
27
Multi-level Ready Queues

• Multiple ready queues
• For different types of processes (e.g., system,
  user)
• For different priority processes (e.g., Mach)
• Each queue can
• Have a different scheduling algorithm
• Receive a different amount of CPU time
• Have movement of processes to another queue
  (feedback)
• e.g., if a process uses too much CPU time, put in
  a lower priority queue
• If a process is getting too little CPU time, put
  it in a higher priority queue

28
Figure 5.6 Multilevel Queue Scheduling
29
Multi-level Feedback Queues (Fig. 5.7)
30
Multiprocessor Scheduling

• Load sharing
• Scheduling problem becomes more complex
• Types of ready queues
• Local dispatch to a specific processor
• Global dispatch to any processor (load
  sharing)
• Processor/process relationship
• Run on only a specific processor (e.g., if it
  must use a device on that processors private
  bus)
• Run on any processor
• Symmetric Each processor does own scheduling
• Master/slave
• Master processor dispatches processes to slaves

31
Symmetric Multiprocessing Synchronization Issues

• Involves synchronization of access to global
  ready queue
• E.g., only one processor must execute a job at
  one time
• Processors CPU1, CPU2, CPU3,
• When a processor (e.g., CPU1) accesses the ready
  queue
• If they attempt access to the ready queue, all
  other processors (CPU2, CPU3, ) must wait
  denied access
• Accessing processor (e.g., CPU1) removes a
  process from ready queue, and dispatchs process
  on itself
• Just before dispatch, that processor makes ready
  queue again available for use by the other CPUs
  (CPU2, CPU3, )

32
Pre-emptive Scheduling Data Structures

• With pre-emptive CPU scheduling, a new process
  can run when an interrupt occurs
• Consider the following situation
• a) Process A in middle of updating data
  structures (in files or RAM), and due to an
  interrupt, is put into ready queue
• b) CPU schedules process B to run
• c) Process B accesses data structures that were
  being modified by Process A
• Problem Data structure update was not completed
  by Process A data structures are in an
  inconsistent state
• Need mechanisms to support cooperative data
  access
• When one thread is updating data structures, we
  need to ensure that it completes the updates
  before other threads access the data structure
• These mechanisms need to work even with
  pre-emptive scheduling!

33
Preemptive vs. Non-Preemptive Kernels (Chapter 6
p. 238-229 p. 256-257)

• Kernels can be preemptive or non-preemptive
• E.g., when a process is running a system call, a
  hardware timer may or may not be able to change
  the running process
• Prior to the Linux 2.6 kernel, Linux was
  non-preemptive
• Linux 2.6 (stable kernel version is 2.6.28.7 as
  of 3/6/09 http//www.kernel.org) is preemptive,
  but uses spinlocks (SMP), and on uni-processors
  will disable kernel premption in some cases