Chapter 5: Process Scheduling - 國立臺灣大學mll.csie.ntu.edu.tw/course/os_f08/lecture/lecture05.pdf · – FCFS, SJF, RR, Multilevel queue, Multilevel feedback queue ... (SRTF)

1

Chapter 5: Process SchedulingChapter 5: Process Scheduling

Lecture 5 

10 14 2008

1

10.14.2008

Hao‐Hua Chu

Announcement

• Nachos assignment #2 out on course webpageNachos assignment #2 out on course webpage– Due in two weeks

– TA will talk about this assignment today.

2

2

Persuasive computing

• Using computing to• Using computing to change human behaviors– Baby Think It Over

– Textrix VR Bike

– Slot Machine

• How to encourage good eating habit in young children?

Example: Playful Tray (NTU)

• Sense to recognize behavior– Weight sensor underneath the tray to 

sense eating actions 

– Eating actions as game input

• Play to engage behavior change– Interactive games: coloring cartoon 

character or penguin fishing

4

character or penguin fishing

3

Outline

• The scheduling problemThe scheduling problem

• Defining  evaluation criteria– CPU utilization, throughput, turnaround time, waiting time, 

response time

• Scheduling algorithms– FCFS, SJF, RR, Multilevel queue, Multilevel feedback queue

– Preemptive vs Non‐preemptivePreemptive vs. Non preemptive

• Evaluating scheduling algorithms– Deterministic model, queueing model, simulation, implementation

• Others– Multiple‐Processor scheduling, process affinity, load balancing, 

thread Scheduling 5

The Scheduling Problem for the CPU Scheduler

• How to maximize CPU utilization in multiprogramming?How to maximize CPU utilization in multiprogramming?– Multiprogramming

– CPU utilization – other performance parameters?

• Consider that a process follows the CPU–I/O Burst Cycle– Process execution consists of a cycle of CPU execution and I/O 

wait

– A typical CPU burst distribution (next slides): Web browser, web yp ( )server

6

4

Alternating sequence of CPU And I/O bursts

7

Histogram of CPU‐burst Times

8

5

Problem solving skill

• Before designing your solutions ….Before designing your solutions ….

• Model the problem– Put the problem in a form that can be solved (using an algorithm)

• Define the evaluation metrics– “Measure” how successful is a solution

• Now design your solutions

9

Modeling the Scheduling Problem

• For the CPU schedulerFor the CPU scheduler– Selects from among the processes in memory (ready queue) that 

are ready to execute, and allocates the CPU to one of them.

• When does the CPU scheduler make a scheduling decision?

10

6

CPU Scheduling Decisions

• CPU scheduling decisions may take place when a process:CPU scheduling decisions may take place when a process:1. Switches from running to waiting state: how does it happen?

2. Switches from running to ready state

3. Switches from waiting to ready

4. Terminates

• Scheduling under 1 and 4 is nonpreemptive

• All other scheduling is preemptive• All other scheduling is preemptive

• Preemption:  (chinese definition)

11

Dispatcher

• Dispatcher module gives control of the CPU to the processDispatcher module gives control of the CPU to the process selected by the short‐term scheduler; this involves:– switching context

– switching to user mode

– jumping to the proper location in the user program to restart that program

• Dispatch latency – time it takes for the dispatcher to stop p y p pone process and start another running

12

7

Scheduling Criteria

• CPU utilization – keep the CPU as busy as possibleCPU utilization  keep the CPU as busy as possible

• Throughput – # of processes that complete their execution per time unit

• Turnaround time – amount of time to execute a particular process

• Waiting time – amount of time a process has been waiting i th din the ready queue

• Response time – amount of time it takes from when a request was submitted until the first response is produced, not output  (for time‐sharing environment)

13

More Scheduling Criteria

• Max CPU utilizationMax CPU utilization

• Max throughput

• Min turnaround time 

• Min waiting time 

• Min response time

P P P

Response time

For P2

Submitted time

P1 P2 P3

24 27 300

Waiting time

Turnaround time

14

8

First‐Come, First‐Served (FCFS) SchedulingProcess Burst Time

P1 24

P2 3

P3 3

• Suppose that the processes arrive in the order: P1 , P2 , P3  The Gantt Chart for the schedule is:

• Waiting time for P1 = 0; P2 = 24; P3 = 27

• Average waiting time:  (0 + 24 + 27)/3 = 17

P1 P2 P3

24 27 300

15

FCFS Scheduling (Cont.)

Suppose that the processes arrive in the orderSuppose that the processes arrive in the order

P2 , P3 , P1• The Gantt chart for the schedule is:

P1P3P2

63 300• Waiting time for P1 = 6; P2 = 0; P3 = 3

• Average waiting time:   (6 + 0 + 3)/3 = 3

• Why is this much better than previous case?– Convoy effect short process behind long process

16

9

Shortest‐Job‐First (SJR) Scheduling

• Associate with each process the length of its next CPUAssociate with each process the length of its next CPU burst.  Use these lengths to schedule the process with the shortest time

• Two schemes: – Nonpreemptive: once CPU given to the process it cannot be 

preempted until completes its CPU burst

– Preemptive: if a new process arrives with CPU burst length less p p gthan remaining time of current executing process, preempt.  This scheme is know as the Shortest‐Remaining‐Time‐First (SRTF)

• SJF is optimal: give minimum average waiting time for a given set of processes

17

Example of Non‐Preemptive SJF

Process Arrival Time Burst TimeProcess Arrival Time Burst Time

P1 0.0 7

P2 2.0 4

P3 4.0 1

P4 5.0 4

• SJF (non‐preemptive)

• Average waiting time = (0 + 6 + 3 + 7)/4  = 4

P1 P3 P2

73 160

P4

8 12

18

10

Example of Preemptive SJF

Process Arrival Time Burst TimeProcess Arrival Time Burst Time

P1 0.0 7

P2 2.0 4

P3 4.0 1

P4 5.0 4

• SJF (preemptive)

5

2

0

0

0

0

• Average waiting time = (9 + 1 + 0 +2)/4 = 3

P1 P3P2

42 110

P4

5 7

P2 P1

16

19

Shortest‐Job‐First (SJR) is Optimal

• SJF is optimalSJF is optimal– Give minimum average waiting time for a given set of processes

• What is the strong assumption in SJF?– Know CPU burst time a‐priori

• How to estimate (guess) the CPU burst time?

• Is SJR still optimal if the estimated CPU burst time is i t?incorrect?

• What is the other assumption in SJF?– All processes have equal importance to users.

20

11

Determining Length of Next CPU Burst

• Can only estimate the lengthCan only estimate the length

• Can be done by using the length of previous CPU bursts, using exponential averaging

10 , 3.

burst CPU next the for value predicted 2.

burst CPU of lenght actual 1.

1

n

thn nt

:Define 4.

.1 1 nnn t

21

Prediction of the Length of the Next CPU Burst

.11 nnn t

50 5.0

22

12

Examples of Exponential Averaging

• =0 .11 nnn t

– n+1 = n– Recent history does not count

• =1– n+1 = tn– Only the actual last CPU burst counts

• If we expand the formula, we get:

1 nnn

n+1 =  tn+(1 ‐ ) tn‐1 + …

+(1 ‐ )j tn ‐j + …

+(1 ‐ )n +1 0

• Since both  and (1 ‐ ) are less than or equal to 1, each successive term has less weight than its predecessor

23

Priority Scheduling

• A priority number (integer) is associated with eachA priority number (integer) is associated with each process

• The CPU is allocated to the process with the highest priority (smallest integer = highest priority)– Non‐preemptive

– Preemptive ‐> when …

• SJF is actually a priority scheduling algorithm• SJF is actually a priority scheduling algorithm– What is the used priority?  

– The predicted next CPU burst time

24

13

Priority Scheduling

• What are the potential problems with priority scheduling?What are the potential problems with priority scheduling?– Can a process starve (although it is ready but never had a chance 

to be executed)?

– Starvation – low priority processes may never execute

• What to solve starvation?– Aging – as time progresses increase the priority of the process

• Priority scheduling is unfair to all processes– How to create a (preemptive) scheduling algorithm that is fair to 

all processes? 

– That is, share the processor fairly among all processes

25

Round Robin (RR)

• Each process gets a small unit of CPU time (timeEach process gets a small unit of CPU time (time quantum), usually 10‐100 milliseconds.  – After this time has elapsed, the process is preempted and added 

to the end of the ready queue.

• If there are n processes in the ready queue and the time quantum is q, then each process gets 1/n of the CPU time in chunks of at most q time units at once.  q– What is the maximum waiting time for a process? 

– (n‐1)q time units

26

14

Example of RR with Time Quantum = 20

Process Burst Time

P1 53

P2 17

P3 68

P4 24

• The Gantt chart is: 

33

48

4

13

28

0

0

8

• How does preemptive RR compare with preemptive SJF?– Average turnaround time? Average waiting time? 

P1 P2 P3 P4 P1 P3 P4 P1 P3 P3

0 20 37 57 77 97 117 121 134 154 162

27

Round Robin (RR)

• How does the size of time quanta (q) affect performance?How does the size of time quanta (q) affect performance?– Effects on context switching time & turnaround time?

– Large q?

• q large  FIFO, smaller context switching time

– Small q?

• q small  larger context switching time, turnaround time may increase.

F l 3 f 10 ti it h 1 ( t d• For example: 3 processes of 10 time units each.  q=1 (avg turnaround time = 29) vs. q=10 (avg turnaround time = 20)

28

15

Effect of Time Quanta Size on Context Switch Time

• Smaller time quantum size leads to more context switchesSmaller time quantum size leads to more context switches

29

Effect of Time Quanta Size on Turnaround Time

• Smaller timeSmaller time quanta “may” increase turnaround time. 

30

16

Round Robin & Shortest‐Job‐First

• Their strong assumption is that all processes have equalTheir strong assumption is that all processes have equal importance to users.– background (batch ‐ virus checking or video compressing) vs. 

foreground processes (interactive processes ‐ word document). 

• Foreground processes: waiting/response time is important (excessive context switching is okay).

• Background processes: waiting time is not important (excessive context switching decreases throughput)context switching decreases throughput)

• What is the solution?– Is it possible to have multiple scheduling algorithms (running at 

the same time) for different types of processes?

31

Multilevel Queue

• Ready queue is partitioned into separate queues:Ready queue is partitioned into separate queues:– foreground (interactive)

– background (batch)

• Each queue has its own scheduling algorithm– foreground – RR, smaller time quantum 

– background – FCFS, larger time quantum

• Scheduling must be done between the queues• Scheduling must be done between the queues– Fixed priority scheduling: serve all from foreground then from 

background.

– Time slice: each queue gets a certain amount of CPU time

• 80% to foreground in RR and 20% to background in FCFS 

32

17

Multilevel Queue Scheduling

33

Multilevel Queue

• Multilevel queue assumes processes are permanentlyMultilevel queue assumes processes are permanently assigned to a queue at the start.

• Two problems: not flexible & starvation– It is difficult to predict the nature of a process at the start (the 

level of interactivity, CPU‐bound, I/O‐bound, etc.) 

– In the fixed‐priority scheduling, low priority process may never be executed. 

• Good thing: low overhead

• How to solve these two problems with a bit more overhead?

34

18

Multilevel Feedback Queue

• A process can move between the various queues; agingA process can move between the various queues; aging can be implemented this way

• Multilevel‐feedback‐queue scheduler defined by the following parameters:– number of queues

– scheduling algorithms for each queue

– method used to determine when to upgrade a processmethod used to determine when to upgrade a process

– method used to determine when to demote a process

– method used to determine which queue a process will enter when that process needs service

35

Example of Multilevel Feedback Queue

• Feedback: based on the observed process’s CPU burst,Feedback: based on the observed process s CPU burst, enables upgrade or downgrade between queues. – CPU‐bound process downgrades to low priority queue.

– I/O‐bound process upgrades to high priority queue. 

• Three queues: – Q0 – RR with time quantum 8 milliseconds

– Q1 – RR time quantum 16 millisecondsQ1 RR time quantum 16 milliseconds

– Q2 – FCFS

36

19

Example of Multilevel Feedback Queue

• SchedulingScheduling– A new job enters queue Q0 which is served FCFS. When it gains 

CPU, job receives 8 milliseconds.  If it does not finish in 8 milliseconds, job is moved to queue Q1.

– At Q1 job is again served FCFS and receives 16 additional milliseconds.  If it still does not complete, it is preempted and moved to queue Q2.

• Comparison to Multilevel Queue:• Comparison to Multilevel Queue:– Benefits: flexibility & no starvation

– Cost: scheduling overhead  

37

Multilevel Feedback Queues

38

20

Multiple‐Processor Scheduling

• Scheduling more complex for multi‐processor systemsScheduling more complex for multi processor systems– Asymmetric multiprocessing vs. symmetric multiprocessing (SMP)

• Asymmetric: master vs. slave processor, only the master processor controls CPU scheduler

• Symmetric: each processor has its own CPU scheduler

• Tradeoff: synchronization overhead (running queue) vs. concurrency

– Processor Affinity

• A higher cost in process migration to another processor why?• A higher cost in process migration to another processor, why?

– A process’s cache memory on a specific processor

– Load Balancing

• Keep load evenly distributed on processors

• Counteracts with processor affinity

– Hyperthreading – logical CPUs, similar to physical CPUs39

Thread Scheduling

• Two scopes of schedulingTwo scopes of scheduling– Global Scheduling: the CPU scheduler decides which kernel thread 

to run next

– Local Scheduling: the user‐level threads library decides which thread to put onto an available LWP

40

21

How to evaluate scheduling algorithms? 

• Deterministic modeling – takes a particularDeterministic modeling  takes a particular predetermined workload and defines the performance of each algorithm  for that workload

• Queueing models – analytical model based on probability– Distributions on process arrival time, CPU burst, I/O time, 

completion time, tec.

• SimulationSimulation– Randomly generated CPU load or collect real traces

• Implementation

41

Pthread Scheduling API

#include <pthread.h>#include <stdio h>#include <stdio.h>#define NUM THREADS 5int main(int argc, char *argv[]){

int i;pthread t tid[NUM THREADS];pthread attr t attr;/* get the default attributes */pthread attr init(&attr);/* set the scheduling algorithm to PROCESS or SYSTEM */pthread attr setscope(&attr, PTHREAD SCOPE SYSTEM);/* set the scheduling policy - FIFO, RT, or OTHER */pthread attr setschedpolicy(&attr, SCHED OTHER);/* create the threads */for (i = 0; i < NUM THREADS; i++)

pthread create(&tid[i],&attr,runner,NULL);

42

22

Pthread Scheduling API/* now join on each thread */

for (i = 0; i < NUM THREADS; i++)

pthread join(tid[i], NULL);

}

/* Each thread will begin control in this function */

void *runner(void *param)

{

printf("I am a thread\n");

pthread exit(0);

}

43

Operating System Examples

• Solaris schedulingSolaris scheduling

• Windows XP scheduling

• Linux scheduling

44

23

Solaris 2 Scheduling

45

Solaris Dispatch Table 

• Multilevel feedback queue for the Interactive and time‐Multilevel feedback queue for the Interactive and timesharing class

46

24

Windows XP Priorities

47

Linux Scheduling

• Two algorithms: time‐sharing and real‐timeg g

• Time‐sharing– Prioritized credit‐based – process with most credits is scheduled 

next

– Credit subtracted when timer interrupt occurs

– When credit = 0, another process chosen

– When all processes have credit = 0, recrediting occurs• Based on factors including priority and history• Based on factors including priority and history

• Real‐time– Soft real‐time

– Posix.1b compliant – two classes• FCFS and RR

• Highest priority process always runs first48

25

The Relationship Between Priorities and Time‐slice length

49

List of Tasks Indexed According to Prorities

50

26

Review

• The scheduling problemThe scheduling problem

• Defining  evaluation criteria– CPU utilization, throughput, turnaround time, waiting time, 

response time

• Scheduling algorithms– FCFS, SJF, RR, Multilevel queue, Multilevel feedback queue

– Preemptive vs Non‐preemptivePreemptive vs. Non preemptive

• Evaluating scheduling algorithms– Deterministic model, queueing model, simulation, implementation

• Others– Multiple‐Processor scheduling, process affinity, load balancing, 

thread Scheduling 51

End of Chapter 5

52