Operating Systems CSE 411 CPU Management Sept. 20 2006 - Lecture 7 Instructor: Bhuvan Urgaonkar.

Operating SystemsOperating SystemsCSE 411CSE 411

CPU ManagementCPU Management

Sept. 20 2006 - Lecture 7Sept. 20 2006 - Lecture 7

Instructor: Bhuvan UrgaonkarInstructor: Bhuvan Urgaonkar

• Last class– Dispatcher– Context switch, FCFS, SJF, SRPT, RR, Priority,

Quantum size– Read on your own: Multi-level (feedback) queue

based scheduling

• Today– More on CPU scheduling

Proportional-Share Schedulers

• A general class of scheduling algorithms• Process Pi given a CPU weight wi > 0• The scheduler needs to ensure the following

– forall i, j, |Ti(t1, t2)/Tj(t1,t2) - wi/wj| ≤ e– Given Pi and Pj were backlogged during [t1,t2]

• Who chooses the weights and how?• Application modeling problem: non-trivial

– Approaches: analytical, empirical• A part of my Ph.D. thesis

• Many PS schedulers developed in the 90s– E.g., Start-time Fair Queueing (Qlinux

UT-Austin/Umass-Amherst)

Lottery Scheduling[Carl Waldspurger, MIT,

~1995]• Perhaps the simplest proportional-share scheduler• Create lottery tickets equal to the sum of the weights

of all processes– What if the weights are non-integral?

• Draw a lottery ticket and schedule the process that owns that ticket– What if the process is not ready?

• Draw tickets only for ready processes

– Homework 1: Calculate the time/space complexity of the operations Lottery scheduling will involve

– Likely question on Exam 1 !!!

Lottery Scheduling Example

1 4

2 5

3 6

7 10

8 11

9 12

13

14

15

9

P1=6 P2=9

Schedule P2

Lottery Scheduling Example

1 4

2 5

3 6

7 10

8 11

9 12

13

14

15

3

P1=6 P2=9

Schedule P1

Lottery Scheduling Example

1 4

2 5 11

6

7 10

8

3 9 12

13

14

15

11

• As t ∞, processes will get their share (unless they were blocked a lot)• Problem with Lottery scheduling: Only probabilistic guarantee• What does the scheduler have to do

– When a new process arrives?– When a process terminates?

P1=6 P2=9

Schedule P2

Work Conservation• Examples of work-conserving schedulers:

All schedulers we have studied so far• Examples of non-work-conserving

schedulers: – DFS, a PS scheduler for multi-processors (Abhishek

Chandra, Umass, 2000, now at Univ. of Minnesota)• Experiments showed the scheduler had become NWC!• Fair Airport to convert into a WC scheme

NWCScheduler

WCScheduler

Reservation-based Schedulers

• Each process has a pair (x, y) – Divide time into periods of length y

each– Guaranteed to get x time units every

period

• Can be Non-work-conserving

Rate Regulation: Leaky Bucket

• A type of reservation-based scheduling• Leaky bucket policing

– Rate ri for process Pi

– CPU cycles over period t ri * t

ri

bi

CPU requirement

CPU the process gets

Max. rate at which the process may progress

burst

Rate Regulation: Token Bucket

• A type of reservation-based scheduling• Token bucket policing

– Rate ri and burst bi for process Pi

– CPU cycles over period t ri * t + bi

ri

bi tokensCPU requirementburst

Deadline-based Scheduling

• Can be NWC• Several variants NP-hard• Real-time systems• “Soft” real-time systems

– E.g., media servers: 30 MPEG-1 frames/sec

– A few violations may be tolerable

An Interesting Problem: An Invitation to do Research on

CPU Scheduling

• Can we achieve the effect of a scheduler that can provide resource guarantees using a priority-based scheduler?

• You are welcome to talk to me during office hours if you find this interesting or if you have any thoughts on this

Hierarchical Schedulers

• Variety of schedulers• Subsets of processes with

different scheduling needs

Reservation-based

(4, 10) (6, 10)

LotteryRound-robinw=1 w=2

UNIXUNIXProcesses

Scheduler Considerations: Context-Switch

Overhead• Switching context

– User mode to Kernel mode– Save PCB– Process interrupt (e.g, TCP/IP processing) if needed– Run scheduling algorithm– Load registers (e.g., PC) from PCB of chosen process – Flush TLB (will discuss this under memory management)

• Switching to user mode• Jumping to the proper location in the user program to restart

that program• Context switch time for Pentium/Linux: tens of microsec• So how often should the scheduler be invoked?

Scheduler Considerations:

Quantum Length• The duration for which a CPU-intensive process will

run before being scheduled out in favor of another process

• Different processes may have different quantum lengths– E.g., UNIX: Higher priority => Larger quantum

• Typically tens of msec in modern systems• Small => Context switch overhead high• Large => Bad responsiveness => Interactive

processes suffer

Scheduler Considerations:CPU Accounting

• Scheduler maintains CPU usage, last scheduling instant etc. in each PCB

• Who should be charged for the CPU usage during interrupt processing?– Optional reading: Resource containers (Rice

University, OSDI 1999, Gaurav Banga et al.)

Scheduler Considerations:Time and Space Requirements

• Run time (n processes)– FCFS: O(1)– RR: O(1)– Deadline-based algos: NP-hard variants, poly-time

heuristics

• Update time: Operations done when set of processes changes (new, terminate, block, become ready)

• Space requirements– Space to store various data structs

Scheduler Evaluation• Optimize one or more of response time,

waiting time, throughput, fairness, utilization, ..

• Overheads– Run time (time to pick the next process to schedule)– Update time– Space requirements

Analytical Approaches to Scheduler Evaluation

• Example 1: Our evaluation of FCFS and SJF• Example 2: Queueing Theory

– Little’s Law: applies to any scheduling discipline– avg_num_processes_in_system = tput *

avg_time_spent_in_system

– Queuing theory provides expressions for mean response time for certain scheduling policies (FCFS, PS, LCFS)

– Other statistical properties of response time only under restrictive assumptions on arrival process and service time

runningwaiting

yet to arrive serviced

Discussion: Analytical Approaches

• Pros:– Intellectually satisfying!– Quick, no code to be written– May provide generally applicable results

• Cons:– Often over-simplified

• Hard to model complex scheduling algorithms• Lack of info about workloads

Simulation-based Evaluation

• Write software that mimics the behavior of the scheduler

• Subject it to realistic workload• Observe output and make inferences about

performance of scheduler• Only need to know the algorithms, no need to

do complex math• Easier to implement than the actual system

Prototyping• Build it!

Analysis, Simulation, and Prototyping

Compared

Analysis

Simulation

Prototyping

accuracy,programming skills

effort, time investment,mathematical skills, generality

Hybrid Approaches

Analysis

Simulation Prototyping

Threads

What is a Thread?

• A basic unit of CPU utilization like a process (not necessarily known to the OS though)

• “Smaller” than a process– Part of a process– Shares code + data + some other OS resources with

other threads that belong to the same process

User Threads• Thread management done by user-level

threads library• OS doesn’t know about the existence of these

threads• Three primary thread libraries:

– POSIX Pthreads– Win32 threads– Java threads

Kernel Threads• OS sees and manages these threads• OS provides system calls to create, terminate,

etc. (just like the system calls it provides for processes)

• Examples– Windows XP/2000– Solaris– Linux– Tru64 UNIX– Mac OS X

Benefits• Responsiveness• Resource Sharing• Economy• Utilization of MP Architectures