Process Scheduling
Basic Question: Which process goes next?
• Personal Computers
– Few processes, interactive, low response time
• Batch Systems (Compute Servers)
– Many processes, not interactive, throughput is
important
• Real Time Systems
– Guaranteed response times, must meet deadlines
Cpr E 308 Spring 2005
Context Switch
• One process stops and another one starts
• Process Contexts stored in the Kernel
Process Table
• Significant Overhead for each Context
Switch
Cpr E 308 Spring 2005
Non Preemptive Scheduling
• Context Switch happens only when the
process gives up the CPU
(wait for I/O, process completes, etc)
– Example: DOS
• Bad: A user process can cause the system to
hang
• Good: Fewer context switches
Cpr E 308 Spring 2005
Preemptive Scheduling
• CPU forcibly taken away from process and
given to a new process
• Preemptive Scheduling used on current PC
operating systems (Windows, Linux)
• Enabled by interrupts
Cpr E 308 Spring 2005
Goals of Batch Scheduling
•
•
•
•
Maximize Throughput (jobs per minute)
Fast Response Times
Fairness to Processes
Often the goals conflict with each other…
Cpr E 308 Spring 2005
First Come First Serve (FCFS)
(non preemptive)
•
Rule: Service processes (jobs) according to order of arrival
•
Implementation
1. Maintain queue of current processes
2. Schedule first in queue; Scheduled job executes until it finishes or blocks
for I/O
3. A new job is added to the end of the queue
•
Advantages: Simple, easy to implement
•
Problems
–
–
–
Example: a short job arriving soon after a long one
What about simultaneous, or near simultaneous arrivals?
Doesn’t optimize throughput or the response time
Cpr E 308 Spring 2005
Shortest Job First
(non-preemptive)
• Suppose 3 jobs
arrive
8
2
1
• Shortest Job First:
3
11
Sum of response times = 15
1
• Scheduler X:
8
Cpr E 308 Spring 2005
10 11
Sum of response times = 29
Shortest Remaining Time Next (preemptive)
• Shortest job First is fine if all jobs arrive
simultaneously
• If not, use shortest remaining time next
– Strategy: Next schedule the job with the
shortest time to completion
• Starvation: A long job might be continually
pre-empted by many short jobs
Cpr E 308 Spring 2005
Realistically . . .
• Don’t know the length of jobs
• Jobs don’t run to completion
– run till they block for user input
• How long is that?
– predict using past behavior
– Guess next job’s running time = average
running time of all jobs so far from the user
Cpr E 308 Spring 2005
Copyright © 2002 Thomas W.Doeppner. All rights reserved
Mean Guessing
Ti
guess(i)
i
Cpr E 308 Spring 2005
Copyright © 2002 Thomas W.Doeppner. All rights reserved
Exponential Averaging
• guessi = a(new_data) + (1-a)(guessi-1)
– a: weight given to new data
– (1-a): weight given to previous guess
Cpr E 308 Spring 2005
Copyright © 2002 Thomas W.Doeppner. All rights reserved
Better Guessing
Ti
guess(i)
i
Cpr E 308 Spring 2005
Copyright © 2002 Thomas W.Doeppner. All rights reserved
Round Robin (Preemptive)
•
•
•
•
Give CPU time to each process by turn
Circular queue of ready processes
Quantum = CPU time allotted every turn
How to choose the value of the quantum?
o Too small might mean high context switch overhead
o Too large might mean bad response times
o Typically a few 10s of milliseconds
o If quantum is very high, then similar to FCFS
Cpr E 308 Spring 2005
Round Robin Example
• Three processes
– P1 = 24ms
– P2 = 6ms
– P3 = 3ms
• Quantum = 4ms
• What is the schedule?
Cpr E 308 Spring 2005
Round Robin
• Better than shortest remaining time next
– Long job will not be overwhelmed by short
jobs
• Better than first come first serve
– Long job arriving early will not overwhelm
future short jobs
Cpr E 308 Spring 2005
Priority Scheduling
• Different Priorities
– Background process sending mail versus
– Shell process accepting input
• “nice” command in UNIX
– allows user to reduce the priority of a process
Cpr E 308 Spring 2005
Priority Scheduling Goals
• Schedule jobs with highest priority first
• High priority jobs should not overwhelm
the others
– Might decrease priority with time
Cpr E 308 Spring 2005
Multiple Queues – Static Version
System level processes
High
Interactive processes
Batch processes
Low
• Lower priority queues don’t run if higher priority queues non-empty
• Could time slice between queues
Cpr E 308 Spring 2005
Multiple Queues – Dynamic Version
Quantum = 1
High
Quantum = 2
Quantum = 4
Low
• Process enters high priority level
• If takes more than 1 quantum, moved to second highest
priority, …..
Cpr E 308 Spring 2005
Multiple Queues – Dynamic Version
• Higher priority = faster service, but short service
times
• Reduce number of context switches for long
processes
• What about a process which became interactive
after a long bout of computation?
– Increase priority?
Cpr E 308 Spring 2005
Real-Time Scheduling
• Provide time guarantees
• Upper bound on response times
– Programmer’s job!
– Every level of the system
• Soft versus hard real-time
– Streaming mp3 player versus air-traffic controller
Cpr E 308 Spring 2005
Real-Time Scheduling
• Is this set of processes schedulable?
– A: once every 100ms, 20ms CPU time
– B: once every 200ms, 80ms CPU time
– C: once every 100ms, 50ms CPU time
• If schedulable, still have to draw up a
schedule for the processes
Cpr E 308 Spring 2005
CPU Scheduling: Summary
• FCFS
• Shortest job first, shortest remaining job next
• Round robin – context switch overhead
• Priority scheduling – Multi-level queues
• Real-time Scheduling – Schedulability
Cpr E 308 Spring 2005