Carnegie Mellon

Scheduling

Some of the slides are adapted from Matt Welsh’s.

Some slides are from Tanenbaum, Modern Operating Systems 3 e, (c) 2008 Prentice-Hall, Inc. All rights
reserved. 0-13-6006639
Some slides are from Silberschatz, and Gagne.
 Carnegie Mellon

Scheduling
 Have already discussed context switching
 Have not discussed how the OS decides which thread to run next
 Context switching is the mechanism
 Scheduling is the policy
 Which thread to run next?
 How long does it run for (granularity)?
 How to ensure every thread gets a chance to run
(fairness)?
 How to prevent starvation?
 Carnegie Mellon

Scheduler
 The scheduler is the OS component that determines which
thread to run next on the CPU
 The scheduler operates on the ready queue
 Why does it not deal with the waiting thread queues?
 When does the scheduler run?
 Carnegie Mellon

Scheduler
 The scheduler is the OS component that determines which thread to run next on
the CPU
 The scheduler operates on the ready queue
 Why does it not deal with the waiting thread queues?
 When does the scheduler run?
 When a thread voluntarily gives up the CPU (yield)
 When a thread blocks on I/O, timer, etc.
 When a thread exits
 When a thread is preempted (e.g., due to timer interrupt)
 Scheduling can be preemptive or non-preemptive
 Preemptive: Timer interrupt can force context switch
 Non-preemptive: Process must yield or block voluntarily
 Batch vs. Interactive Scheduling
 Batch: Non-preemptive and no other jobs run if they block
 Interactive: Preemptive and other jobs do run if they block
 Carnegie Mellon

Dispatcher
 Dispatcher 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

one process and start another running
 Carnegie Mellon

Scheduling Policy Goals

 Goal of a scheduling policy is to achieve some “optimal”
allocation of CPU time in the system
 According to some definition of “optimal”
 Possible goals of the scheduling policy??
 Carnegie Mellon
Note: The term jobs (left from
the early days of OS) can be
Scheduling criteria - 1 replaced by ”processes”.

 CPU utilization: The percentage of time that CPU is running (user) jobs
 Throughput: The number of jobs completed per unit time
 120 jobs completed over 1 minute:
 Throughput: 2 jobs/second
 Turnaround time: The duration from the submission of the job to its completion.
 Job submitted at 8:05:00, and completed at 8:15:00.
 Turnaround time: 10:00
 Carnegie Mellon
Note: The term jobs (left from
the early days of OS) can be
Scheduling criteria - 2 replaced by ”processes”.

 Waiting time: The total amount of time a job spent in the ready list.
 Load average: The average number of jobs in the ready queue.
 Response time: The amount of time a request is submitted until the
first response is produced.
 The amount of time from pressing a key on the keyboard, and
seeing it printed on the screen.
 Carnegie Mellon

Scheduling Policy Goals

 Possible goals:
 Maximize CPU utilization
 Maximize throughput
 Minimize turnaround time
 Minimize response time
 Minimize waiting time
 These goals often conflict!
 Batch system: Try to maximize job throughput and minimize
turnaround time
 Interactive system: Minimize response time of interactive jobs (i.e.,
editors, etc.)
 The choice of scheduling policy has a huge impact on
performance
 Carnegie Mellon

Starvation
 Schedulers often try to eliminate thread starvation
 e.g., If a high priority thread always gets to run before a low-
priority thread
 We say the low priority thread is starved
 Not all schedulers have this as a goal!
 Sometimes starvation is permitted in order to achieve other goals
 Example: Real time systems
 Some threads must run under a specific deadline
 e.g., Motor-control task must run every 30 ms to effectively steer
robot
 In this case it is (sometimes) OK to starve other threads
 Carnegie Mellon

First-Come-First-Served (FCFS)
 Jobs are scheduled in the order that they arrive
 Also called First-In-First-Out (FIFO)
 Used only for batch scheduling
 Implies that job runs to completion – never blocks or gets context switched
out
 Jobs treated equally, no starvation!
 As long as jobs eventually complete, of course
 What's wrong with FCFS?

Job A Job B Job C

time
Short jobs get stuck behind long ones!
 Carnegie Mellon

Round Robin (RR)

 Essentially FCFS with preemption
 A thread runs until it blocks or its CPU quantum expires
 How to determine the ideal CPU quantum?
FCFS Job A Job B Job C
time

RR
time
 Job A: 13 time units, Job B & C: 4 time units
 Turnaround time with FCFS: Job A = 13, Job B = (13+4), Job C = (13 + 4 + 4)
Total turnaround time = 51, mean = (51/3) = 17
 Turnaround time with RR: Job A = 21, Job B = 11, Job C = 12
 Total turnaround time = 44, mean = (44/3) = 14.667
 Carnegie Mellon

Shortest Job First (SJF)

 Schedule job with the shortest expected CPU burst
 Two broad classes of processes: CPU bound and I/O
bound
 CPU bound:
cpu i/o cpu i/o cpu i/o

 I/O bound:
cpu i/o cpu i/o cpu i/o cpu

 Examples of each kind of process?

 Carnegie Mellon

Shortest Job First (SJF)

 Schedule job with the shortest expected CPU burst
 Two broad classes of processes: CPU bound and I/O
bound
 CPU bound: e.g.compiler, number crunching, games.
cpu i/o cpu i/o cpu i/o

 I/O bound: e.g. web browser, database engine, word processor.

cpu i/o cpu i/o cpu i/o cpu

 How to predict a process's CPU burst?

 Can get a pretty good guess by looking at the past history of the job
 Track the CPU burst each time a thread runs, track the average
 CPU bound jobs will tend to have a long burst
 I/O bound jobs will tend to have a short burst
 SJF Example
Carnegie Mellon

Job A cpu i/o

Job B cpu i/o

Job C cpu i/o

Resulting schedule:

B i/o

A i/o

B i/o

A i/o B is not on the ready queue!

C i/o

B i/o
 Carnegie Mellon

Shortest Job First (SJF)

 Schedule job with the shortest expected CPU burst
 This policy is non-preemptive. Job will run until it blocks for I/O.
 SJF scheduling prefers I/O bound processes. Why?
 Idea: A long CPU burst “hogs” the CPU.
 Running short-CPU-burst jobs first gets them done, and out of the way.
 Allows their I/O to overlap with each other: more efficient use of the CPU
 Interactive programs often have a short CPU burst: Good to run them first
 To yield “snappy” interactive performance, e.g., for window system or
shell.
 We all do this. It is called “procrastination.”
 When faced with too much work, easier to do the short tasks first, get
them out of the way.
 Leave the big, hard tasks for later.
 Carnegie Mellon

Shortest Remaining Time First (SRTF)

 SJF is a nonpreemptive policy.
 Preemptive variant: Shortest Remaining Time First (SRTF)
• If a job becomes runnable with a shorter expected CPU burst,
preempt current job and run the new job
B i/o

A Preempt A when B becomes runnable

B i/o

A i/o When A becomes runnable C is not

preempted and SRT_A > SRT_C

B i/o

C i/o
 SRTF versus RR
Carnegie Mellon

 Say we have three jobs:

 Job A and B: both CPU-bound, will run for hours on the CPU with no I/O
 Job C: Requires a 1ms burst of CPU followed by 10ms I/O operation
 RR with 25 ms time slice: C C
A B A

Job C's I/O

 RR with 1 ms time slice:

Job C's I/O

 Lots of pointless context switches between Jobs A and B!

 SRTF:

 Job A runs to completion, then Job B starts

 Carnegie Mellon

Comparison of FCFS, RR, SJF and SRTF

FCFS RR SJF SRTF

Preemptive? N Y N Y

When is the scheduler called? FCFS RR SJF SRTF

Current process exits Y Y Y Y
Current process goes for I/O N Y Y Y
A new process is added N N N Y
Timer interrupt goes off N Y N N
A process returns from I/O - N N Y
 Carnegie Mellon

Priority Scheduling
 Assign each thread a priority
 In Linux, these range from 0 (lowest) to 99 (highest)
 UNIX “nice()” system call lets user adjust this
 But note, scale is inverted: -20 is highest priority and +20 is lowest
 Priority may be set by user, OS, or some combination of the two
 User may adjust priority to bias scheduler towards a thread
 OS may adjust priority to achieve system performance goals
 When scheduling, simply run the job with the highest priority
 Usually implemented as separate “priority queues”
 One queue for each priority level
 Use RR scheduling within each queue
 If a queue is empty, look in next lowest priority queue
 Problem: Starvation
 High priority threads always trump low priority threads
 Carnegie Mellon

Multilevel Feedback Queues (MLFQ)

 Observation: Want to give higher priority to I/O-bound jobs
 They are likely to be interactive and need CPU rapidly after I/O completes
 However, jobs are not always I/O bound or CPU-bound during execution!
 Web browser is mostly I/O bound and interactive
 But, becomes CPU bound when running a Java applet
 Basic idea: Adjust priority of a thread in response to its CPU
usage
 Increase priority if job has a short CPU burst
 Decrease priority if job has a long CPU burst (e.g., uses up CPU quantum)
 Jobs with lower priorities get longer CPU quantum
 What is the rationale for this???
 Carnegie Mellon

Multilevel Feedback Queues (MLFQ)

 Observation: Want to give higher priority to I/O-bound jobs
 They are likely to be interactive and need CPU rapidly after I/O completes
 However, jobs are not always I/O bound or CPU-bound during execution!
 Web browser is mostly I/O bound and interactive
 But, becomes CPU bound when running a Java applet
 Basic idea: Adjust priority of a thread in response to its CPU usage
 Increase priority if job has a short CPU burst
 Decrease priority if job has a long CPU burst (e.g., uses up CPU quantum)
 Jobs with lower priorities get longer CPU quantum
 What is the rationale for this???
 Don't want to give high priority to CPU-bound jobs...
 Because lower-priority jobs can't preempt them if they get the CPU.
 OK to give longer CPU quantum to low-priority jobs:
 I/O bound jobs with higher priority can still preempt when they become runnable.
 MLFQ Implementation
Carnegie Mellon

Run PID 4277, T0 PID 4391, T2

State: Ready State: Ready
High prio PC PC

Registers Registers

PID 3202, T1
State: Ready
Medium prio PC

Registers

Low prio
 MLFQ Implementation
Carnegie Mellon

PID 4391, T2
State: Ready
High prio PC

Registers Uses entire CPU burst (preempted)

Placed into lower priority queue

PID 3202, T1 PID 4277, T0

State: Ready State: Ready
Medium prio PC PC

Registers Registers

Low prio
 MLFQ Implementation
Carnegie Mellon

Run PID 4391, T2

State: Ready
High prio PC

Registers

PID 3202, T1 PID 4277, T0

State: Ready State: Ready
Medium prio PC PC

Registers Registers

Low prio
 MLFQ Implementation
Carnegie Mellon

High prio

Preempted

PID 3202, T1 PID 4277, T0 PID 4391, T2

State: Ready State: Ready State: Ready
Medium prio PC PC PC

Registers Registers Registers

Low prio
 MLFQ Implementation
Carnegie Mellon

High prio

PID 3202, T1 PID 4277, T0 PID 4391, T2

Run State: Ready State: Ready State: Ready
Medium prio PC PC PC

Registers Registers Registers

Low prio
 MLFQ Implementation
Carnegie Mellon

PID 3202, T1 Runs with short CPU burst

State: Ready (blocks on I/O)
High prio PC

Registers

PID 4277, T0 PID 4391, T2

State: Ready State: Ready
Medium prio PC PC

Registers Registers

Low prio
 Carnegie Mellon

Priority inversion
 A problem that may occur in priority scheduling systems!
 A high priority process is indirectly ”preempted” by a
lower priority task effectively "inverting" the relative
priorities of the two tasks.
 It happened on the Mars rover Sojourner!

http://www.drdobbs.com/jvm/what-is-priority-inversion-and-how-do-yo/230600008
https://users.cs.duke.edu/~carla/mars.html
 Carnegie Mellon

Priority inversion
A gets in the B gets B’s I/O B gets
ready queue A blocks on in ends and in
resource R A acquires lock
and I/O preempts C I/O for resource R
preempts C
and runs
High A A

Medium B B
Low C C C

C acquires B gets in the B runs C

a lock for ready queue C runs B runs releases
resource R but waits for lock
A

B ”seems to” have a higher

priority than A!
Hence priority inversion!
 Carnegie Mellon

Solution: Priority inheritance

A gets in the A acquires lock
ready queue A blocks on for resource R
and preempts C resource R and runs
High A A

Medium
Low C C

C acquires C runs with

a lock for A’s priority
resource R and releases
lock

C “inherits” A’s priority!

Hence priority inheritance!

http://www.drdobbs.com/jvm/what-is-priority-inversion-and-how-do-yo/230600008
https://users.cs.duke.edu/~carla/mars.html
 Carnegie Mellon

Lottery Scheduling
 A kind of randomized priority scheduling scheme!
 Give each thread some number of “tickets”
 The more tickets a thread has, the higher its priority
 On each scheduling interval:
 Pick a random number between 1 and total # of tickets
 Scheduling the job holding the ticket with this number
 How does this avoid starvation?
 Even low priority threads have a small chance of running!
 Carnegie Mellon

Lottery scheduling example

Job A Job B Job C

0-30 30-40 40-99

30 tickets 10 tickets 60 tickets

Round 1 26 A i/o

Round 2 65 C i/o

Round 3 92 C would win ... but it is still blocked!

Round 4 33 B i/o

Round 5 7 A i/o
 Carnegie Mellon

Guaranteed Scheduling
 Provide guarantees about CPU usage
 If there are N processes, then each should get 1/N of CPU
allocation.
 How to do it?
 For each process
 Compute the ratio of actual CPU time / consumed CPU time.
 Pick the one with the lowest ratio.
 Ratio of 0.5: process had consumed half of it should have had
 Ratio of 2.0: process had consumed twice of it should have had
 Carnegie Mellon

Fair-share scheduling
 We have assumed that each process is of its own, with no
regard who its owner is.
 CPU allocation is split to the number of processes a user
has.
 A user running a single process would run 10 times as fast,
than another user running 10 copies of the same process.
 Carnegie Mellon

Multi-Processor Scheduling
 CPU scheduling more complex when multiple CPUs are
available

 Homogeneous processors within a multiprocessor system

 multiple physical processors
 single physical processor providing multiple logical processors
 hyperthreading
 multiple cores
 Carnegie Mellon

Multiprocessor scheduling
 On a uniprocessor:
 Which thread should be run next?
 On a multiprocessor:
 Which thread should be run on which CPU next?
 What should be the scheduling unit?
 Threads or processes
 Recall user-level and kernel-level threads
 In some systems all threads are independent,
 Independent users start independent processes
 in others they come in groups
 Make
 Originally compiles sequentially
 Newer versions starts compilations in parallel
 The compilation processes need to be treated as a group and scheduled
to maximize performance
 Carnegie Mellon

Multi-Processor Scheduling
 Asymmetric multiprocessing
 A single processor (master) handles all the scheduling with regard to CPU, I/O for all the
processors in the system.
 Other processors execute only user code.
 only one processor accesses the system data structures, alleviating the need for data sharing
 Symmetric multiprocessing (SMP)
 Two or more identical processors are connected to a single shared main memory.
 Most common multiprocessor systems today use an SMP architecture
 Each processor does his own self-scheduling.
 Carnegie Mellon

Issues with SMP scheduling - 1

 Processor affinity
 Migration of a process from one processor to another is costly
 cached data is invalidated
 Avoid migration of one process from one processor to another.
 Hard affinity: Assign a processor to a particular process and do not
allow it to migrate.
 Soft affinity: The OS tries to keep a process running on the same
processor as much as possible.
 http://www.linuxjournal.com/article/6799
 Carnegie Mellon

Issues with SMP scheduling - 2

 Load balancing
 All processors should keep an eye on their load with respect to the
load of other processors
 Processes should migrate from loaded processors to idle ones.
 Push migration: The busy processor tries to unload some of its
processes
 Pull migration: The idle process tries to grab processes from other
processors
 Push and pull migration can run concurrently
 Load balancing conflicts with processor affinity.
 Space sharing
 Try to run threads from the same process on different CPUs
simultaneously
 Carnegie Mellon

Real-Time Scheduling
 Hard real-time systems
 System must always complete a
critical task within a guaranteed
amount of time
 On-board computer system of a
robot
 Designers must describe task
requirements
 Worst-case execution time of
instruction sequences
 “Prove” system response time
 Carnegie Mellon

Real-Time Scheduling

 Soft real-time systems

 requires that critical processes receive
priority over less fortunate ones
 Carnegie Mellon

On-line demo
 http://sehitoglu.web.tr/scheddemo/