Last Time

1
Last Time

• Processes
• Mechanisms Dispatcher
• while (1) stop A, store A, load B, run B ?
  Easy!
• But, not so straightforward
• How does the dispatcher gain control of the
  processor?
• Saving context, registers really tricky?
• Whos B? ? Today

2
CPU Scheduling
UNIVERSITY of WISCONSIN-MADISONComputer Sciences
Department
CS 537Introduction to Operating Systems
Andrea C. Arpaci-DusseauRemzi H.
Arpaci-Dusseau Haryadi S. Gunawi

• Questions answered in this lecture
• What is scheduling vs. allocation?
• What is preemptive vs. non-preemptive scheduling?
• What are FCFS, SJF, STCF, RR?
• What are their advantages and disadvantages?

3
Types of Resources

• Resources can be classified into one of two
  groups
• 1) Preemptible resources
• Can take resource away, give it back later
• Resource has little state associated with it
• May only have one of this resource
• Example CPU, memory, i/o channel
• 2) Non-preemptible resources
• Once given resource, cannot be reused until
  voluntarily relinquished
• Need many instances of this resource
• Example Blocks on disk
• Cant we make disk space preemptible?
• Yes, if we have backup tapes. But since, disk is
  the last level of storage in our PCs, disk space
  is not preemptible

4
OS Management

• Type of resource determines how
• the OS manages it
• Non-preemptible allocation policy
• Decide which process gets which resource
• Ex store a 1-GB file, the OS (specifically file
  system) will allocate the space
• Preemptible scheduling policy
• Decide order in which requests are serviced
• Decide how long process keeps resource
• Ex CPU scheduling (this lecture)
• Ex Memory swapping (put most recently-used data
  in the memory, and put least-recently-used data
  in the disk)

5
Levels of CPU Management

• Dispatcher
• Low-level mechanism
• Performs context-switch
• Scheduler
• Policy to determine which process gets CPU when
• Allocator
• Policy to determine which processes compete for
  which CPU
• Needed for multiprocessor, parallel, and
  distributed systems

6
CPU Workload Model

• Workload contains collection of jobs (processes)
• Job model
• Job alternates between CPU and I/O bursts (i.e.,
  moves between ready and blocked queues)
• CPU-bound job Long CPU bursts
• I/O-bound job Short CPU bursts

7
3 states of a processWho controls which
transition
Running
Ready
Blocked

• Running ? Blocked
• controlled by the application (e.g. could only
  happen if the application calls an I/O operation,
  or waiting for some events to happen)
• Blocked ? Ready
• controlled by the I/O or event mechanism.
• Ready ? Running
• controlled by the scheduler
• Running ? Ready
• controlled by the scheduler, could only happen
  if theres a hardware interrupt

8
Non-preemptive vs. preemptive scheduler

• Although CPU is a preemptible resource, we can
  design a non-preemptive scheduler or a preemptive
  scheduler
• Non-preemptive scheduler
• Process remains scheduled until voluntarily
  relinquishes CPU
• Preemptive scheduler
• Process may be descheduled at any time

9
Scheduling Performance Metrics

• Minimize waiting time
• Do not want to spend much time in Ready queue
• (In this course, we simplify the definition of
  wait time. Wait time the difference between the
  time a process arrives and the time the process
  gets the CPU for the first time)
• Minimize turnaround time
• Do not want to wait long for job to complete
• Maximize throughput
• Want many jobs to complete per unit of time
• Maximize resource utilization
• Keep expensive devices busy (CPU, disks)
• How can CPU scheduling affects I/O (e.g. disk)
  utilization??? See 2 examples in this notes
• Minimize response time
• Schedule interactive jobs promptly so users see
  output quickly
• Minimize overhead
• Reduce number of context switches
• Maximize fairness
• All jobs get same amount of CPU over some time
  interval

10
Gantt Chart

• Illustrates how jobs are scheduled over time on
  CPUExample

11
First-Come-First-Served (FCFS)

• Idea Maintain FIFO list of jobs as they arrive
• Non-preemptive policy
• Allocate CPU to job at head of list

Average wait time(0 (10-1)
(12-2))/36.33 Average turnaround time (10
(12-1) (16-2))/311.67
12
FCFS Discussion

• Advantage Very simple implementation
• Disadvantage
• Waiting time depends on arrival order
• Potentially long wait for jobs that arrive later
• Convoy effect Short jobs stuck waiting for long
  jobs
• Hurts waiting time of short jobs
• Reduces utilization of I/O devices
• Example 1 mostly CPU-bound job (process A), 3
  mostly I/O-bound jobs (process B, C, D)
• Process A uses CPU a lot, and calls I/O rarely

CPU
Disk
Idle
Time
13
Advanced FIFOPreemptive vs. non-preemptive FIFO

• FIFO with I/Os
• In the last 2 slides, two examples were given for
  FCFS. The processes in the first example perform
  no I/O (hence FIFO looks simple!).
• In the second one, the processes perform some
  I/Os.
• Preemptive or no-preemptive FIFO?
• Thus, if a process A is currently running and
  then performs an I/O, it will go to the blocked
  queues. After the I/O finishes, A is ready, but
  the process B is hugging the CPU. The question is
  should A take over the CPU or not?
• The answer depends whether you want to create a
  preemptive or non-preemptive FIFO.
• Example
• Process A, at time 1-10 uses the CPU, at time
  11-15 uses the disk, and needs to continue for
  another 5 CPU time
• Process B needs 10 CPU time, and it comes in
  later than A
• Gantt Chart (see next slide)

14
Preemptive vs. non-preemptive FIFO

• Non-preemptive FIFO will look like this
• Why? Because at t15, the scheduler cannot
  preempt B, thus A must wait in the ready queue
• Thus, FIFO does not look like a FIFO since B is
  out first than A

A
B
A
25
0
10
20
Time

• Preemptive FIFO will look like this
• Preemptive means the scheduler can take over the
  CPU anytime it wants
• In this case, my preemptive scheduler policy
  specifies that if a process (A) in a ready queue
  arrived earlier than the arrival time of running
  process (B). The earlier process (A) should take
  over the CPU.

A
B
A
B
15
25
0
10
20
Time
15
Shortest-Job-First (SJF)

• Idea Minimize average wait time by running
  shortest CPU-burst next
• Minimize convoy effect
• Use FCFS if jobs are of same length
• Non-preemptive SJF looks like this

Average wait (0 2 6) / 3 Average
turnaround (2 6 16) / 3
A
B
C
0
2
6
16
Time
16
SJF Discussion

• Advantages
• Provably optimal for minimizing average wait time
  (with no preemption)
• Moving shorter job before longer job improves
  waiting time of short job more than it harms
  waiting time of long job
• Helps keep I/O devices busy
• Disadvantages
• Not practical Cannot predict future CPU burst
  time
• OS solution Use past behavior to predict future
  behavior (i.e. adaptive priority, next lecture)
• Is this a stupid policy? No, 40 years ago, in
  batch programming system, SJF was literally used
  in real systems.
• Starvation Long jobs may never be scheduled

17
Non-preemptive vs. Preemptive SJF

• Given the arrival time like described in the
  table below, show the Gantt chart for
  non-preemptive and preemptive SJF. The preemptive
  one is shown below. How about for non-preemptive?

If you think about this for a while, preemptive
SJF doesnt make sense because the scheduler
always uses the initial CPU burst for making a
decision. However, some of the CPU burst may have
been served. Thus, preemptive SJF should use the
remaining CPU burst for making the decision.
Hence, STCF comes into play (see next slide)
A
B
C
A
B
0
1
7
2
4
16
Time
18
Shortest-Time-to-Completion-First (STCF or SCTF)

• Idea Add preemption to SJF
• Schedule newly ready job if shorter than
  remaining burst for running job

Non-preemptive SJF Average wait (0 8 12
17) / 4 STCF Average wait (0 1 5 17) / 4 ?
win a lot!
? Non-preemptive SJF
12
26
17
0
8

• STCF (see
• next slide for
• more details)

26
17
10
1
5
0
Time
19
STCF (timing details)

• Colored CPU time means the process is being served

26
17
10
1
5
0
20
STCF

• Advantages
• Better than SJF in terms of wait time
• Similar to SJF, overall turnaround time is much
  better than FIFO (the idea is give the most to
  those that need the least)
• Disadvantages same with SJF
• Not practical, no future information
• Starvation (solution? Round-robin ? next slide)

21
Round-Robin (RR)

• Idea Run each job for a time-slice (quantum) and
  then move to back of FIFO queue
• Preempt job if still running at end of time-slice
• Example quantum 1 CPU unit

Average wait (0 0 0) / 3 Turnaround time
(16 (5-1) (10-2)) / 3
A
B
C
A
B
C
A
C
A
C
A
Time
22
RR timing details
How about if quantum 2 unit?
23
RR Discussion

• Advantages
• Jobs get fair share of CPU
• Shortest jobs finish relatively quickly (compare
  the turnaround time of RR and FIFO using the
  previous example)
• Disadvantages
• Poor average waiting time with similar job
  lengths
• Example 10 jobs each requiring 10 time slices,
  and quantum 1
• RR All complete after about 100 time slices
• P1 ends at 91, P2 at 92, P3 at 93, and so on
• FCFS performs better!
• P1 ends at 10, P2 at 20, P3 at 30, P4 at 40
• Is this another stupid policy?
• No, in the old days, only 2-3 people working on
  the machine, and only small number of processes
• Performance depends on length of time-slice
• If time-slice too short, pay overhead of context
  switch
• If time-slice too long, degenerate to FCFS
• How to choose a good quantum?

24
RR Time-Slice

• IF time-slice too long, degenerate to FCFS
• Example
• Job A w/ 1 ms compute and 10ms I/O
• Job B always computes
• Time-slice is 100 ms ? underutilize I/O ! Disk
  utilization is about 10

CPU
B
A
B
A
Disk
A
Idle
A
Idle
Time
25
RR Time-Slice (contd)

• Goal Adjust length of time-slice to match I/O
  burst
• In this example, e.g. 10 ms
• We want something like this
• I/O is not underutilized! Disk utilization is
  high
• Lots of overlapping between CPU and I/O
  utilization ? good

CPU
B
A
B
A
B
A
B
A
Disk
A
A
A
A
Time
26
Summary

• So far SJF, STCF, Round-Robin all bad
• Solution priority-based? (next lecture)
• Recap
• No best general-purpose scheduling policy
• All depends on the workload
• Need to know which workload fits with which
  scheduler
• Rule of thumbs
• History and future are important
• Unfortunately future information not always
  available
• Hence, predict the future (next lecture)