Thread Scheduling

Goals for Today

• Scheduling Policy goals

• Policy Options
• Implementation Considerations

Note: Some slides and/or pictures in the following are adapted from slides ©2005
Silberschatz, Galvin, and Gagne. Slides courtesy of Anthony D. Joseph, John
Kubiatowicz, AJ Shankar, George Necula, Alex Aiken, Eric Brewer, Ras Bodik,
Ion Stoica, Doug Tygar, and David Wagner.
 CPU Scheduling

• Earlier, we talked about the life-cycle of a thread

– Active threads work their way from Ready queue to Running to
various waiting queues.
• Question: How is the OS to decide which of several threads to
take off a queue?
– Obvious queue to worry about is ready queue
– Others can be scheduled as well, however
• Scheduling: deciding which threads are given access to resources
 Scheduling Assumptions
• CPU scheduling big area of research in early 70’s
• Many implicit assumptions for CPU scheduling:
– One program per user
– One thread per program
– Programs are independent
• In general unrealistic but they simplify the problem
– For instance: is “fair” about fairness among users or programs?
» If I run one compilation job and you run five, you get five times as
much CPU on many operating systems
• The high-level goal: Dole out CPU time to optimize some desired
parameters of system

USER USER2 USER USER1 USE

1 3 R2
Time
 Assumption: CPU Bursts

Weighted toward small bursts

• Execution model: programs alternate between bursts of CPU and

I/O
– Program typically uses the CPU for some period of time, then does
I/O, then uses CPU again
– Each scheduling decision is about which job to give to the CPU for
use by its next CPU burst
– With timeslicing, thread may be forced to give up CPU before
finishing current CPU burst
 Scheduling Metrics
• Waiting Time: time the job is waiting in the ready queue
– Time between job’s arrival in the ready queue and launching the job
• Service (Execution) Time: time the job is running
• Response (Completion) Time:
– Time between job’s arrival in the ready queue and job’s completion
– Response time is what the user sees:
» Time to echo a keystroke in editor
» Time to compile a program

Response Time = Waiting Time + Service Time

• Throughput: number of jobs completed per unit of time

– Throughput related to response time, but not same thing:
» Minimizing response time will lead to more context switching than if
you only maximized throughput
 Scheduling Policy Goals/Criteria

• Minimize Response Time

– Minimize elapsed time to do an operation (or job)

• Maximize Throughput
– Two parts to maximizing throughput
» Minimize overhead (for example, context-switching)
» Efficient use of resources (CPU, disk, memory, etc)

• Fairness
– Share CPU among users in some equitable way
– Fairness is not minimizing average response time:
» Better average response time by making system less fair
 First-Come, First-Served (FCFS) Scheduling
• First-Come, First-Served (FCFS)
– Also “First In, First Out” (FIFO) or “Run until done”
» In early systems, FCFS meant one program
scheduled until done (including I/O)
» Now, means keep CPU until thread blocks
• Example: Process Burst Time
P1 24
P2 3
P3 3
– Suppose processes arrive in the order: P1 , P2 , P3
The Gantt Chart for the schedule is:
P1 P2 P3

0 24 27 30
– Waiting time for P1 = 0; P2 = 24; P3 = 27
– Average waiting time: (0 + 24 + 27)/3 = 17
– Average completion time: (24 + 27 + 30)/3 = 27
• Convoy effect: short process behind long process
 FCFS Scheduling (Cont.)
• Example continued:
– Suppose that processes arrive in order: P2 , P3 , P1
Now, the Gantt chart for the schedule is:
P2 P3 P1

0 3 6 30
– Waiting time for P1 = 6; P2 = 0; P3 = 3
– Average waiting time: (6 + 0 + 3)/3 = 3
– Average Completion time: (3 + 6 + 30)/3 = 13
• In second case:
– Average waiting time is much better (before it was 17)
– Average completion time is better (before it was 27)
• FCFS Pros and Cons:
– Simple (+)
– Short jobs get stuck behind long ones (-)
» Safeway: Getting milk, always stuck behind cart full of small items
 Round Robin (RR)
• FCFS Scheme: Potentially bad for short jobs!
– Depends on submit order
– If you are first in line at supermarket with milk, you don’t care who
is behind you, on the other hand…
• Round Robin Scheme
– Each process gets a small unit of CPU time
(time quantum), usually 10-100 milliseconds
– After quantum expires, the process is preempted
and added to the end of the ready queue
– n processes in ready queue and time quantum is q ⇒
» Each process gets 1/n of the CPU time
» In chunks of at most q time units
» No process waits more than (n-1)q time units
• Performance
– q large ⇒ FCFS
– q small ⇒ Interleaved
– q must be large with respect to context switch, otherwise overhead is
too high (all overhead)
 Example of RR with Time Quantum = 20
• Example: Process Burst Time Remaining Time
P1 53 53
P2 8 8
P3 68 68
P4 24 24
– The Gantt chart is:
 Example of RR with Time Quantum = 20
• Example: Process Burst Time Remaining Time
P1 53 33
P2 8 8
P3 68 68
P4 24 24
– The Gantt chart is:
P1
0 20
 Example of RR with Time Quantum = 20
• Example: Process Burst Time Remaining Time
P1 53 33
P2 8 0
P3 68 68
P4 24 24
– The Gantt chart is:
P1 P2
0 20 28
 Example of RR with Time Quantum = 20
• Example: Process Burst Time Remaining Time
P1 53 33
P2 8 0
P3 68 48
P4 24 24
– The Gantt chart is:
P1 P2 P3
0 20 28 48
 Example of RR with Time Quantum = 20
• Example: Process Burst Time Remaining Time
P1 53 33
P2 8 0
P3 68 48
P4 24 4
– The Gantt chart is:
P1 P2 P3 P4
0 20 28 48 68
 Example of RR with Time Quantum = 20
• Example: Process Burst Time Remaining Time
P1 53 13
P2 8 0
P3 68 48
P4 24 4
– The Gantt chart is:
P1 P2 P3 P4 P1
0 20 28 48 68 88
 Example of RR with Time Quantum = 20
• Example: Process Burst Time Remaining Time
P1 53 13
P2 8 0
P3 68 28
P4 24 4
– The Gantt chart is:
P1 P2 P3 P4 P1 P3
0 20 28 48 68 88 108
 Example of RR with Time Quantum = 20
• Example: Process Burst Time Remaining Time
P1 53 0
P2 8 0
P3 68 0
P4 24 0
– The Gantt chart is:
P1 P2 P3 P4 P1 P3 P4 P1 P3 P3
0 20 28 48 68 88 108 112 125 145 153

– Waiting time for P1=(68-20)+(112-88)=72

P2=(20-0)=20
P3=(28-0)+(88-48)+(125-108)=85
P4=(48-0)+(108-68)=88
– Average waiting time = (72+20+85+88)/4=66¼
– Average completion time = (125+28+153+112)/4 = 104½
• Thus, Round-Robin Pros and Cons:
– Better for short jobs, Fair (+)
– Context-switching time adds up for long jobs (-)
 Round-Robin Discussion
• How do you choose time slice?
– What if too big?
» Response time suffers
– What if infinite (∞)?
» Get back FCFS/FIFO
– What if time slice too small?
» Throughput suffers!
• Actual choices of timeslice:
– Initially, UNIX timeslice one second:
» Worked ok when UNIX was used by one or two people.
» What if three compilations going on? 3 seconds to echo each
keystroke!
– In practice, need to balance short-job performance and long-job
throughput:
» Typical time slice today is between 10ms – 100ms
» Typical context-switching overhead is 0.1ms – 1ms
» Roughly 1% overhead due to context-switching
 Comparisons between FCFS and Round Robin
• Assuming zero-cost context-switching time, is RR always better
than FCFS?
• Simple example: 10 jobs, each takes 100s of CPU time
RR scheduler quantum of 1s
All jobs start at the same time
FCFS P1 P2 … P9 P10
0 100 200 800 900 1000
RR … … … … …
0 10 20 980 990 1000
991 999
• Completion Times: Job # FIFO RR
1 100 991
2 200 992
… … …
9 900 999
10 1000 1000
 Comparisons between FCFS and Round Robin
• Assuming zero-cost context-switching time, is RR always better
than FCFS?
• Simple example: 10 jobs, each takes 100s of CPU time
RR scheduler quantum of 1s
All jobs start at the same time
FCFS P1 P2 … P9 P10
0 100 200 800 900 1000
RR … … … … …
0 10 20 980 990 1000
991 999
• Both RR and FCFS finish at the same time
• Average response time is much worse under RR!
– Bad when all jobs same length
• Also: Cache state must be shared between all jobs with RR but can
be devoted to each job with FCFS
– Total time for RR longer even for zero-cost switch!
 Earlier Example with Different Time Quantum
P2 P4 P1 P3
Best FCFS:
[8] [24] [53] [68]
0 8 32 85 153
Quantum P1 P2 P3 P4 Average
Best FCFS 32 0 85 8 31¼

Wait
Time

Best FCFS 85 8 153 32 69½

Completion
Time
 Earlier Example with Different Time Quantum
P3 P1 P4 P2
Worst FCFS:
[68] [53] [24] [8]
0 68 121 145 153
Quantum P1 P2 P3 P4 Average
Best FCFS 32 0 85 8 31¼

Wait
Time

Worst FCFS 68 145 0 121 83½

Best FCFS 85 8 153 32 69½

Completion
Time

Worst FCFS 121 153 68 145 121¾

 Earlier Example with Different Time Quantum
P3 P1 P4 P2
Worst FCFS:
[68] [53] [24] [8]
0 68 121 145 153
Quantum P1 P2 P3 P4 Average
P1 P2 P3 P4 PBest
1 PFCFS
3 P4 P1 32
P3 P4 P01 P3 P185 P3 P1 8P3 P1 31¼
P3 P3 P3
Q=1 84 22 85 57 62
0 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128 133 141 149 153
Q=5 82 20 85 58 61¼
Wait
Q=8 80 8 85 56 57¼
Time
Q = 10 82 10 85 68 61¼
Q = 20 72 20 85 88 66¼
Worst FCFS 68 145 0 121 83½
Best FCFS 85 8 153 32 69½
Q=1 137 30 153 81 100½
Q=5 135 28 153 82 99½
Completion
Q=8 133 16 153 80 95½
Time
Q = 10 135 18 153 92 99½
Q = 20 125 28 153 112 104½
Worst FCFS 121 153 68 145 121¾
 What if we Knew the Future?
• Could we always mirror best FCFS?
• Shortest Job First (SJF):
– Run whatever job has the least amount of
computation to do

• Shortest Remaining Time First (SRTF):

– Preemptive version of SJF: if job arrives and has a shorter time to
completion than the remaining time on the current job, immediately
preempt CPU

• These can be applied either to a whole program or the current

CPU burst of each program
– Idea is to get short jobs out of the system
– Big effect on short jobs, only small effect on long ones
– Result is better average response time
 Discussion

• SJF/SRTF are the best you can do at minimizing average

response time
– Provably optimal (SJF among non-preemptive, SRTF among
preemptive)
– Since SRTF is always at least as good as SJF, focus on SRTF

• Comparison of SRTF with FCFS and RR

– What if all jobs the same length?
» SRTF becomes the same as FCFS (i.e., FCFS is best can do if all jobs
the same length)
– What if jobs have varying length?
» SRTF (and RR): short jobs not stuck behind long ones
 Example to illustrate benefits of SRTF

A or B C

C’s C’s C’s

I/O I/O I/O
• Three jobs:
– A,B: CPU bound, each run for a week
C: I/O bound, loop 1ms CPU, 9ms disk I/O
– If only one at a time, C uses 90% of the disk, A or B use 100% of the
CPU
• With FIFO:
– Once A or B get in, keep CPU for one week each
• What about RR or SRTF?
– Easier to see with a timeline
 RR vs. SRTF
Disk Utilization:
C A B 9/201 ~ 4.5%
C

C’s RR 100ms time slice Disk

C’s Utilization:
I/O ~90%
I/O but lots of
wakeups!
CABAB… C

RR 1ms time slice

C’s C’s
I/O I/O
Disk Utilization:
C A A A 90%

SRTF
C’s C’s
I/O I/O
 SRTF Further discussion
• Starvation
– SRTF can lead to starvation if many small jobs!
– Large jobs never get to run
• Somehow need to predict future
– How can we do this?
– Some systems ask the user
» When you submit a job, have to say how long it will take
» To stop cheating, system kills job if takes too long
– But: even non-malicious users have trouble predicting runtime of
their jobs
• Bottom line, can’t really know how long job will take
– However, can use SRTF as a yardstick
for measuring other policies
– Optimal, so can’t do any better
• SRTF Pros & Cons
– Optimal (average response time) (+)
– Hard to predict future (-)
– Unfair (-)
 Predicting the Length of the Next CPU Burst
• Adaptive: Changing policy based on past behavior
– CPU scheduling, in virtual memory, in file systems, etc.
– Works because programs have predictable behavior
» If program was I/O bound in past, likely in future
» If computer behavior were random, wouldn’t help
• Example: SRTF with estimated burst length
– Use an estimator function on previous bursts:
Let tn-1, tn-2, tn-3, etc. be previous CPU burst lengths.
Estimate next burst τn = f(tn-1, tn-2, tn-3, …)
– Function f could be one of many different time series estimation
schemes (Kalman filters, etc.)
– Example:
Exponential averaging
τn = αtn-1+(1-α)τn-1
with (0<α≤1)
 Multi-Level Feedback Scheduling

Long-Running
Compute tasks
demoted to
low priority

• Another method for exploiting past behavior

– First used in Cambridge Time Sharing System (CTSS)
– Multiple queues, each with different priority
» Higher priority queues often considered “foreground” tasks
– Each queue has its own scheduling algorithm
» e.g., foreground – RR, background – FCFS
» Sometimes multiple RR priorities with quantum increasing exponentially
(highest:1ms, next:2ms, next: 4ms, etc.)
• Adjust each job’s priority as follows (details vary)
– Job starts in highest priority queue
– If timeout expires, drop one level
– If timeout doesn’t expire, push up one level (or to top)
 Scheduling Details

• Result approximates SRTF:

– CPU bound jobs drop like a rock
– Short-running I/O bound jobs stay near top

• Scheduling must be done between the queues

– Fixed priority scheduling:
» Serve all from highest priority, then next priority, etc.
– Time slice:
» Each queue gets a certain amount of CPU time
» e.g., 70% to highest, 20% next, 10% lowest
 Scheduling Fairness
• What about fairness?
– Strict fixed-priority scheduling between queues is unfair (run highest,
then next, etc):
» Long running jobs may never get CPU
» In Multics, shut down machine, found 10-year-old job
– Must give long-running jobs a fraction of the CPU even when there
are shorter jobs to run
– Tradeoff: fairness gained by hurting average response time!

• How to implement fairness?

– Could give each queue some fraction of the CPU
» What if one long-running job and 100 short-running ones?
» Like express lanes in a supermarket—sometimes express lanes get so
long, get better service by going into one of the other lines
– Could increase priority of jobs that don’t get service
» What is done in UNIX
» This is ad hoc—what rate should you increase priorities?
 Lottery Scheduling
• Yet another alternative: Lottery Scheduling
– Give each job some number of lottery tickets
– On each time slice, randomly pick a winning ticket
– On average, CPU time is proportional to number of tickets given
to each job

• How to assign tickets?

– To approximate SRTF, short running jobs get more, long running
jobs get fewer
– To avoid starvation, every job gets at least one ticket (everyone
makes progress)

• Advantage over strict priority scheduling: behaves gracefully

as load changes
– Adding or deleting a job affects all jobs proportionally,
independent of how many tickets each job possesses
 Lottery Scheduling Example
• Lottery Scheduling Example
– Assume short jobs get 10 tickets, long jobs get 1 ticket

# short jobs/ % of CPU each % of CPU each

# long jobs short jobs gets long jobs gets
1/1 91% 9%
0/2 N/A 50%
2/0 50% N/A
10/1 9.9% 0.99%
1/10 50% 5%

– What if too many short jobs to give reasonable

response time?
» In UNIX, if load average is 100, hard to make progress
» One approach: log some user out
 How to Evaluate a Scheduling algorithm?
• Deterministic modeling
– Takes a predetermined workload and compute the performance of
each algorithm for that workload
• Queuing models
– Mathematical approach for handling stochastic workloads
• Implementation/Simulation:
– Build system which allows actual algorithms to be run against actual
data. Most flexible/general.
 A Final Word On Scheduling
• When do the details of the scheduling policy and fairness really
matter?
– When there aren’t enough resources to go around
• When should you simply buy a faster computer?
– (Or network link, or expanded highway, or …)
– One approach: Buy it when it will pay
for itself in improved response time

Response time
» Assuming you’re paying for worse
response time in reduced productivity,
customer angst, etc…

100%
» Might think that you should buy a
faster X when X is utilized 100%,
but usually, response time goes
to infinity as utilization⇒100% Utilization
• An interesting implication of this curve:
– Most scheduling algorithms work fine in the “linear” portion of the
load curve, fail otherwise
– Argues for buying a faster X when hit “knee” of curve
 Summary
• Scheduling: selecting a process from the ready queue and
allocating the CPU to it

• FCFS Scheduling:
– Run threads to completion in order of submission
– Pros: Simple (+)
– Cons: Short jobs get stuck behind long ones (-)

• Round-Robin Scheduling:
– Give each thread a small amount of CPU time when it executes; cycle
between all ready threads
– Pros: Better for short jobs (+)
– Cons: Poor when jobs are same length (-)
 Summary (cont’d)
• Shortest Job First (SJF)/Shortest Remaining Time First (SRTF):
– Run whatever job has the least amount of computation to do/least
remaining amount of computation to do
– Pros: Optimal (average response time)
– Cons: Hard to predict future, Unfair

• Multi-Level Feedback Scheduling:

– Multiple queues of different priorities
– Automatic promotion/demotion of process priority in order to
approximate SJF/SRTF

• Lottery Scheduling:
– Give each thread a number of tokens (short tasks ⇒ more tokens)
– Reserve a minimum number of tokens for every thread to ensure
forward progress/fairness