1. Chapter 5: CPU Scheduling

2. Chapter 5: Objectives Understand Scheduling Criteria
Scheduling Algorithms
Multiple-Processor Scheduling

3. 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?

4. 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)
Few long bursts
Many short bursts

5. 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:
it has to wait, e.g., for I/O request or for a child to terminate
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)

6. 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?

7. Scheduling Criteria CPU utilization – keep the CPU as busy as possible
Maximize
Throughput – # of processes that complete their execution per time unit
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
Response time – amount of time it takes from when a request was submitted until the first response is produced, not output (for time-sharing environment)

8. 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

9. 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:
Waiting time for P1 = 0; P2 = 24; P3 = 27
Average waiting time: ( )/3 = 17 P1 P2 P3 24 27 30

10. FCFS Scheduling (Cont.)
Suppose that the processes arrive in the order
P2 , P3 , P1
3, 3, 24
The Gantt chart for the schedule is:
Waiting time for P1 = 6; P2 = 0; P3 = 3
Average waiting time: ( )/3 = 3
Much better than previous case
Convoy effect: short process behind long process P1 P3 P2 6 3 30

11. 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

12. Example of Non-Preemptive SJF
Process Arrival Time Burst Time P P P P
SJF (non-preemptive)
Average waiting time = ( )/4 = 4 P1 P3 P2 7 3 16 P4 8 12

13. Example of Preemptive SJF
Process Arrival Time Burst Time P P P P
SJF (preemptive, SRJF)
Average waiting time = ( )/4 = 3 P1 P3 P2 4 2 11 P4 5 7 16

14. 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

15. Exponential Averaging
If we expand the formula, we get:
n+1 =  tn +  (1 - ) tn -1 +  (1 -  )2 tn … + (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)

16. Prediction CPU Burst Lengths: Expo Average
Assume  = 0.5, 0 = 10

17. 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

18. Round Robin (RR)
Each process gets a small unit of CPU time (time quantum), usually 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

19. Example of RR with Time Quantum = 20
Process Burst Time
P1 53
P2 17
P3 68
P4 24
The Gantt chart is:
Typically, higher average turnaround than SJF, but better response P1 P2 P3 P4 20 37 57 77 97
117
121
134
154
162

20. Time Quantum and Context Switch Time
Smaller q  more responsive but more context switches (overhead)

21. 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

22. 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

23. Multilevel Queue Scheduling

24. 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

25. 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

26. Multilevel Feedback Queues
Notes:
Short processes get served faster (higher prio)  more responsive
Long processes (CPU bound) sink to bottom  served FCFS  more throughput

27. 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

28. 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

29. 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

30. Operating System Examples
Windows XP scheduling
Linux scheduling

31. 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

32. 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

33. 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)

34. 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)

35. 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!

36. 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