Module 6: CPU Scheduling - Simon Fraser University

Download Report

Transcript Module 6: CPU Scheduling - Simon Fraser University

Chapter 5: CPU Scheduling
Chapter 5: Objectives
 Understand

Scheduling Criteria

Scheduling Algorithms

Multiple-Processor Scheduling
Operating System Concepts
5.2
Silberschatz, Galvin and Gagne ©2005
Basic Concepts
 Maximum CPU utilization obtained with
multiprogramming
 CPU–I/O Burst Cycle

Process execution consists of a cycle
of CPU execution and I/O wait
 How long is the CPU burst?
Operating System Concepts
5.3
Silberschatz, Galvin and Gagne ©2005
CPU Burst Distribution
 CPU bursts vary greatly from process to process and from computer
to computer
 But, in general, they tend to have the following distribution (expo)
Many short
bursts
Few long bursts
Operating System Concepts
5.4
Silberschatz, Galvin and Gagne ©2005
CPU Scheduler
 Selects one process from ready queue to run on CPU
 Scheduling can be

Nonpreemptive
 Once a process is allocated the CPU, it does not leave unless:
1. it has to wait, e.g., for I/O request or for a child to terminate
2. it terminates
 Preemptive
 OS can force (preempt) a process from CPU at anytime
– Say, to allocate CPU to another higher-priority process
 Which is harder to implement? and why?

Preemptive is harder: Need to maintain consistency of
 data shared between processes, and more importantly,
 kernel data structures (e.g., I/O queues)
– Think of a preemption while kernel is executing a sys call on
behalf of a process (many OSs, wait for sys call to finish)
Operating System Concepts
5.5
Silberschatz, Galvin and Gagne ©2005
Dispatcher
 Scheduler: selects one process to run on CPU
 Dispatcher: allocates CPU to the selected process, which involves:

switching context

switching to user mode

jumping to the proper location (in the selected process) and
restarting it
 Dispatch latency – time it takes for the dispatcher to stop one process
and start another running
 How does scheduler select a process to run?
Operating System Concepts
5.6
Silberschatz, Galvin and Gagne ©2005
Scheduling Criteria
 CPU utilization – keep the CPU as busy as possible

Maximize
 Throughput – # of processes that complete their execution per time unit

Maximize
 Turnaround time – amount of time to execute a particular process (time
from submission to termination)

Minimize
 Waiting time – amount of time a process has been waiting in the ready
queue

Minimize
 Response time – amount of time it takes from when a request was
submitted until the first response is produced, not output (for timesharing environment)

Minimize
Operating System Concepts
5.7
Silberschatz, Galvin and Gagne ©2005
Scheduling Algorithms
 First Come, First Served
 Shortest Job First
 Priority
 Round Robin
 Multilevel queues
 Note: A process may have many CPU bursts, but in the examples
we show only one for simplicity
Operating System Concepts
5.8
Silberschatz, Galvin and Gagne ©2005
First-Come, First-Served (FCFS) Scheduling
Process
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:
P1
P2
0
24
P3
27
30
 Waiting time for P1 = 0; P2 = 24; P3 = 27
 Average waiting time: (0 + 24 + 27)/3 = 17
Operating System Concepts
5.9
Silberschatz, Galvin and Gagne ©2005
FCFS Scheduling (Cont.)
Suppose that the processes arrive in the order
P2 , P3 , P1
3, 3, 24
 The Gantt chart for the schedule is:
P2
0
P3
3
P1
6
30
 Waiting time for P1 = 6; P2 = 0; P3 = 3
 Average waiting time: (6 + 0 + 3)/3 = 3
 Much better than previous case
 Convoy effect: short process behind long process
Operating System Concepts
5.10
Silberschatz, Galvin and Gagne ©2005
Shortest-Job-First (SJF) Scheduling
 Associate 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 than remaining time of current executing process,
preempt. This scheme is know as the
Shortest-Remaining-Time-First (SRTF)
 SJF is optimal – gives minimum average waiting time for a given
set of processes
Operating System Concepts
5.11
Silberschatz, Galvin and Gagne ©2005
Example of Non-Preemptive SJF
Process
Arrival Time
Burst Time
P1
0.0
7
P2
2.0
4
P3
4.0
1
P4
5.0
4
 SJF (non-preemptive)
P1
0
3
P3
7
P2
8
P4
12
16
 Average waiting time = (0 + 6 + 3 + 7)/4 = 4
Operating System Concepts
5.12
Silberschatz, Galvin and Gagne ©2005
Example of Preemptive SJF
Process
Arrival Time
Burst Time
P1
0.0
7
P2
2.0
4
P3
4.0
1
P4
5.0
4
 SJF (preemptive, SRJF)
P1
0
P2
2
P3
4
P2
5
P4
P1
11
7
16
 Average waiting time = (9 + 1 + 0 +2)/4 = 3
Operating System Concepts
5.13
Silberschatz, Galvin and Gagne ©2005
Determining Length of Next CPU Burst
 Can only estimate the length
 Can be done by using the length of previous CPU bursts, using
exponential averaging
1. t n  actual lenght of n th CPU burst
2.  n 1  predicted value for the next CPU burst
3.  , 0    1
4. Define :
 n 1   t n  1   n

Operating System Concepts
5.14
Silberschatz, Galvin and Gagne ©2005
Exponential Averaging
 If we expand the formula, we get:
n+1 =  tn +  (1 - ) tn -1 +  (1 -  )2 tn -2 + … + (1 -  )n +1 0
 Since both  and (1 - ) are less than or equal to 1, each successive term
has less weight than its predecessor
 Examples:

 = 0 ==> n+1 = n ==> Last CPU burst does not count (transient
value)

 =1 ==> n+1 =  tn ==> Only last CPU burst counts (history is stale)
Operating System Concepts
5.15
Silberschatz, Galvin and Gagne ©2005
Prediction CPU Burst Lengths: Expo Average
 Assume  = 0.5, 0 = 10
Operating System Concepts
5.16
Silberschatz, Galvin and Gagne ©2005
Priority Scheduling
 A priority number (integer) is associated with each process
 The CPU is allocated to the process with the highest priority
(smallest integer  highest priority)

Preemptive

nonpreemptive
 SJF is a priority scheduling where priority is the predicted next CPU
burst time
 Problem  Starvation – low priority processes may never execute
 Solution  Aging – as time progresses increase the priority of the
process
 Aging: increase the priority of a process as it waits in the system
Operating System Concepts
5.17
Silberschatz, Galvin and Gagne ©2005
Round Robin (RR)
 Each 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. No process waits more
than (n-1)q time units.
 Performance

q large  FCFS

q small  q must be large with respect to context switch,
otherwise overhead is too high
Operating System Concepts
5.18
Silberschatz, Galvin and Gagne ©2005
Example of RR with Time Quantum = 20
Process
P1
P2
P3
P4
 The Gantt chart is:
P1
0
P2
20
37
P3
Burst Time
53
17
68
24
P4
57
P1
77
P3
97 117
P4
P1
P3
P3
121 134 154 162
 Typically, higher average turnaround than SJF, but better response
Operating System Concepts
5.19
Silberschatz, Galvin and Gagne ©2005
Time Quantum and Context Switch Time
 Smaller q  more responsive but more context switches (overhead)
Operating System Concepts
5.20
Silberschatz, Galvin and Gagne ©2005
Turnaround Time Varies With The Time Quantum
 Turnaround time varies with quantum, then stabilizes
 Rule of thumb for good performance:

80% of CPU bursts should be shorter than time quantum
Operating System Concepts
5.21
Silberschatz, Galvin and Gagne ©2005
Multilevel Queue
 Ready queue is partitioned into separate queues:

foreground (interactive)

background (batch)
 Each queue has its own scheduling algorithm

foreground – RR

background – FCFS
 Scheduling must be done between the queues

Fixed priority scheduling; (i.e., serve all from foreground then from
background). Possibility of starvation.

Time slice – each queue gets a certain amount of CPU time which
it can schedule amongst its processes; e.g.,

80% to foreground in RR

20% to background in FCFS
Operating System Concepts
5.22
Silberschatz, Galvin and Gagne ©2005
Multilevel Queue Scheduling
Operating System Concepts
5.23
Silberschatz, Galvin and Gagne ©2005
Multilevel Feedback Queue
 A process can move between 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 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
Operating System Concepts
5.24
Silberschatz, Galvin and Gagne ©2005
Example of Multilevel Feedback Queue
 Three queues:

Q0 – RR with time quantum 8 milliseconds

Q1 – RR time quantum 16 milliseconds

Q2 – FCFS
 Scheduling

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
Operating System Concepts
5.25
Silberschatz, Galvin and Gagne ©2005
Multilevel Feedback Queues
 Notes:

Short processes get served faster (higher prio)  more responsive

Long processes (CPU bound) sink to bottom  served FCFS 
more throughput
Operating System Concepts
5.26
Silberschatz, Galvin and Gagne ©2005
Multiple-Processor Scheduling
 Multiple processors ==> divide load among them

More complex than single CPU scheduling
 How to divide load?

Asymmetric multiprocessor


One master processor does the scheduling for others
Symmetric multiprocessor (SMP)

Each processor runs its own scheduler

One common ready queue for all processors, or one ready
queue for each

Win XP, Linux, Solaris, Mac OS X support SMP
Operating System Concepts
5.27
Silberschatz, Galvin and Gagne ©2005
SMP Issues

Processor affinity

When a process runs on a processor, some data is cached in that
processor’s cache

A process migrates to another processor ==>





Cache of new processor has to be re-populated
Cache of old processor has to be invalidated
==> performance penalty
Load balancing

One processor has too much load and another is idle

Balance load using

Push migration: A specific task periodically checks load on all
processors and evenly distributes it by moving (pushing) tasks

Pull migration: Idle processor pulls a waiting task from a busy
processor

Some systems (e.g., Linux) implement both
Tradeoff between load balancing and processor affinity: what would you do?

May be, invoke load balancer when imbalance exceeds a threshold
Operating System Concepts
5.28
Silberschatz, Galvin and Gagne ©2005
Real-time Scheduling
 Hard-real time systems

A task must be finished within a deadline

Ex: Control of spacecraft
 Soft-real time systems

A task is given higher priority over others

Ex: Multimedia systems
Operating System Concepts
5.29
Silberschatz, Galvin and Gagne ©2005
Operating System Examples
 Windows XP scheduling
 Linux scheduling
Operating System Concepts
5.30
Silberschatz, Galvin and Gagne ©2005
Windows XP Scheduler
 Priority-based, preemptive scheduler

The highest-priority thread will always run
 32 levels of priorities, each has a separate queue
 Scheduler traverses queues from highest to lowest till it finds a thread
that is ready to run
 Priorities are divided into classes, each has several relative priorities
Operating System Concepts
5.31
Silberschatz, Galvin and Gagne ©2005
Windows XP Scheduler (cont’d)
 Real-time class (fixed): Levels 16 to 31
 Other classes (variable): Levels 1 to 15

Priority may change (decrease or increase)
 Priority decreases

After thread’s quantum time runs out


but never goes below the base (normal) value of its class
Limit CPU consumption of CPU-bound threads
 Priority increases

After a thread is released from a wait operation

Bigger increase if thread was waiting for mouse or keyboard

Moderate increase if it was waiting for disk

Also, active window gets a priority boost

Yield good response time
Operating System Concepts
5.32
Silberschatz, Galvin and Gagne ©2005
Linux Scheduler

Priority-based, preemptive scheduler with two separate ranges

Real-time: 0 to 99

Nice: 100 to 140 (from -20 to 19)

Higher priority tasks get larger quanta (unlike Win XP, Solaris)
Operating System Concepts
5.33
Silberschatz, Galvin and Gagne ©2005
Linux Scheduler (cont’d)
 A task is initially assigned a time slice (quantum)
 Runqueue has two arrays: active and expired
 A runnable task is eligible for CPU if it still has time left in its time slice
 If the time slice runs out, the task is moved to the expired array
 When there are no tasks in the active array, the expired array becomes
the active array and vice versa (change of pointers)
Operating System Concepts
5.34
Silberschatz, Galvin and Gagne ©2005
Algorithm Evaluation
 Deterministic modeling

Takes a particular predetermined workload and defines the
performance of each algorithm for that workload

Not general
 Queuing models

Use queuing theory to analyze algorithms

Many (unrealistic) assumptions to facilitate analysis
 Simulation

Build a simulator and test with

synthetic workload (e.g., generated randomly), or

Traces collected from running systems
 Implementation

Code it up and test!
Operating System Concepts
5.35
Silberschatz, Galvin and Gagne ©2005
Summary
 Process execution: cycle of CPU bursts and I/O bursts

CPU bursts lengths: many short bursts, and few long ones
 Scheduler selects one process from ready queue

Dispatcher performs the switching
 Scheduling criteria (usually conflicting)

CPU utilization, waiting time, response time, throughput, …
 Scheduling Algorithms

FCFS, SJF, Priority, RR, Multilevel Queues, …
 Multiprocessor Scheduling

Processor affinity vs. load balancing
 Evaluation of Algorithms

Modeling, simulation, implementation
 Examples

Win XP, Linux
Operating System Concepts
5.36
Silberschatz, Galvin and Gagne ©2005