Always want to have CPU (or CPU

1

• Always want to have CPU (or CPUs) working
• Usually many processes in ready queue
• Ready to run on CPU
• Focus on a single CPU here
• Need strategies for
• Allocating CPU time
• Selecting next process to run
• Deciding what happens after a process completes a
  system call, or completes I/O
• Short-term scheduling
• Must not take much CPU time to do the scheduling

2
SchedulingIntroduction to Scheduling

• Bursts of CPU usage alternate with periods of I/O
  wait
• a CPU-bound process
• an I/O bound process

3
Scheduling Performance Criteria

• CPU utilization
• Percentage of time that CPU is busy (and not
  idle), over some period of time
• Throughput
• Number of jobs completed per unit time
• Turnaround time
• Time interval from submission of a process until
  completion of the process
• Waiting time
• Sum of the time periods spent in the ready queue
• Response time
• Time from submission until first output/input
• Approximate (estimate) by time from submission
  until first access to CPU

4
CPU Scheduling Algorithms

• First-Come, First-Served (FCFS)
• Complete the jobs in order of arrival
• Shortest Job First (SJF)
• Complete the job with shortest next CPU burst
• Priority (PRI)
• Processes have a priority
• Allocate CPU to process with highest priority
• Round-Robin (RR)
• Each process gets a small unit of time on CPU
  (time quantum or time slice)

5
Solution Gantt Chart Method
FCFS First-Come First-Served

• Waiting times?
• P1 0
• P2 20
• P3 32
• P4 40
• P5 56
• Average wait time 148/5 29.6

6
FCFS First-Come First-Served

• Advantage Relatively simple algorithm
• Disadvantage long waiting times

7
SJF Shortest Job First

• The job with the shortest next CPU burst time is
  selected
• Example (from before)
• CPU job burst times and ready queue order
• P1 20
• P2 12
• P3 8
• P4 16
• P5 4
• Draw Gantt chart and compute the average waiting
  time given SJF CPU scheduling
• Assume 0 context switch time

8
SJF Solution

• Waiting times (how long did process wait before
  being scheduled on the CPU?)
• P1 40
• P2 12
• P3 4
• P4 24
• P5 0
• Average wait time 16

(Recall FCFS scheduling had average wait time of
29.6)
9
SJF

• Provably shortest average wait time
• BUT What do we need to actually implement this?

10
Priority Scheduling

• Have to decide on a numbering scheme
• 0 can be highest or lowest
• Priorities can be
• Internal
• Set according to O/S factors (e.g., memory
  requirements)
• External
• Set as a user policy e.g., User importance
• Static
• Fixed for the duration of the process
• Dynamic
• Changing during processing
• E.g., as a function of amount of CPU usage, or
  length of time waiting (a solution to starvation)

11
Starvation Problem

• Priority scheduling algorithms can suffer from
  starvation (indefinite waiting for CPU access)
• In a heavily loaded system, a steady stream of
  higher-priority processes can result in a low
  priority process never receiving CPU time
• I.e., it can starve for CPU time
• One solution aging
• Gradually increasing the priority of a process
  that waits for a long time

See also Mogul, J. C. and Ramakrishnan, K. K.
(1997). Eliminating receive livelock in an
interrupt-driven kernel. ACM Trans. Comput. Syst.
15, 3 (Aug. 1997), 217-252. DOI
http//doi.acm.org/10.1145/263326.263335
12
Which Scheduling Algorithms Can be Preemptive?

• FCFS (First-come, First-Served)
• Non-preemptive
• SJF (Shortest Job First)
• Can be either
• Choice when a new (shorter) job arrives
• Can preempt current job or not
• Priority
• Can be either
• Choice when a processes priority changes or when
  a higher priority process arrives

13
RR (Round Robin) Scheduling

• Now talking about time-sharing or multi-tasking
  system
• typical kind of scheduling algorithm in a
  contemporary general purpose operating system
• Method
• Give each process a unit of time (time slice,
  quantum) of execution on CPU
• Then move to next process in ready queue
• Continue until all processes completed
• Necessarily preemptive
• Requires use of timer interrupt
• Time quantum typically between 10 and 100
  milliseconds
• Linux default appears to be 100ms

14
RR (Round Robin) Scheduling Example

• CPU job burst times order in queue
• P1 20
• P2 12
• P3 8
• P4 16
• P5 4
• Draw Gantt chart, and compute average wait time
• Time quantum of 4
• Like our previous examples, assume 0 context
  switch time

15
Solution

• Waiting times
• P1 60 - 20 40
• P2 44 - 12 32
• P3 32 - 8 24
• P4 56 - 16 40
• P5 20 - 4 16
• Average wait time 30.4

16
Other Performance Criteria

• Response time
• Estimate by time from job submission to time to
  first CPU dispatch
• Assume all jobs submitted at same time, in order
  given
• Turnaround time
• Time interval from submission of a process until
  completion of the process
• Assume all jobs submitted at same time

17
Response Time Calculations
Job FCFS SJF RR
P1 0 40 0
P2 20 12 4
P3 32 4 8
P4 40 24 12
P5 56 0 16
Average 29.6 16 8
18
Turnaround Time Calculations
Job FCFS SJF RR
P1 20 60 60
P2 32 24 44
P3 40 12 32
P4 56 40 56
P5 60 4 20
Average 41.6 28 42.4
19
Performance Characteristics of Scheduling
Algorithms

• Different scheduling algorithms will have
  different performance characteristics
• RR (Round Robin)
• Average waiting time often high
• Good average response time
• Important for interactive or timesharing systems
• SJF
• Best average waiting time
• Some overhead with respect to estimates of CPU
  burst length ordering ready queue

20
Context Switching Issues

• This analysis has not taken context switch
  duration into account
• In general, the context switch will take time
• Just like the CPU burst of a process takes time
• Response time, wait time etc. will be affected by
  context switch time
• RR (Round Robin) quantum duration
• To reduce response times, want smaller time
  quantum
• But, smaller time quantum increases system
  overhead

21
Example

• Calculate average waiting time for RR (round
  robin) scheduling, for
• Processes
• P1 24
• P2 4
• P3 4
• Assume above order in ready queue P1 at head of
  ready queue
• Quantum 4 context switch time 1

22
Solution Average Wait Time With Context Switch
Time

• P1 0 11 4 15
• P2 5
• P310
• Average 10

(This is also a reason to dynamically vary the
time quantum. E.g., Linux 2.5 scheduler, and Mach
O/S.)
23
Multi-level Ready Queues

• Multiple ready queues
• For different types of processes (e.g., system,
  user)
• For different priority processes (e.g., Mach)
• Each queue can
• Have a different scheduling algorithm
• Receive a different amount of CPU time
• Have movement of processes to another queue
  (feedback)
• e.g., if a process uses too much CPU time, put in
  a lower priority queue
• If a process is getting too little CPU time, put
  it in a higher priority queue

24
Multilevel Queue Scheduling
25
Multi-level Feedback Queues
26
Scheduling in Existing Systems Linux 2.5 kernel

• Priority-based, preemptive
• Two priority ranges (real time and nice)
• Time quantum longer for higher priority processes
  (ranges from 10ms to 200ms)
• Tasks are runnable while they have time remaining
  in their time quantum once exhausted, must wait
  until others have exhausted their time quantum

27
O(1) Background

• Briefly the scheduler maintained two runqueues
  for each CPU, with a priority linked list for
  each priority level (140 total).
• Tasks are enqueued into the corresponding
  priority list.
• The scheduler only needs to look at the highest
  priority list to schedule the next task.
• Assigns timeslices for each task.
• Had to track sleep times, process interactivity,
  etc.

28
O(1) Background

• Two runqueues per CPU, I said...one active, one
  expired. If a process hasn't used its entire
  timeslice, it's on the active queue if it has,
  it's expired. Tasks are swapped between the two
  as needed.
• Timeslice and priority are recalculated when a
  task is swapped.
• If the active queue is empty, they swap pointers,
  so the empty one is now the expired queue.

29
O(1) Background

• The first 100 priority lists are for real-time
  tasks, the last 40 are for user tasks.
• User tasks can have their priorities dynamically
  adjusted, based on their dependency. (I/O or CPU)

30
CPU Scheduling as of Linux 2.6.23 Kernel
Completely Fair Scheduler

• Goal fairness in dividing processor time to
  tasks
• Balanced (red-black) tree to implement a ready
  queue O(log n) insert or delete time
• Queue ordered in terms of virtual run time
• smallest value picked for using CPU
• small values tasks have received less time on
  CPU
• tasks blocked on I/O have smaller values
• execution time on CPU added to value
• priorities cause different decays of values

http//www.ibm.com/developerworks/linux/library/l-
completely-fair-scheduler/
31
The Completely Fair Scheduler

• CFS cuts out a lot of the things previous
  versions tracked no timeslices, no sleep time
  tracking, no process type identification...
• Instead, CFS tries to model an ideal, precise
  multitasking CPU one that could run multiple
  processes simultaneously, giving each equal
  processing power.
• Obviously, this is purely theoretical, so how can
  we model it?

32
CFS, continued

• We may not be able to have one CPU run things
  simultaneously, but we can measure how much
  runtime each task has had and try and ensure that
  everyone gets their fair share of time.
• This is held in the vruntime variable for each
  task, and is recorded at the nanosecond level. A
  lower vruntime indicates that the task has had
  less time to compute, and therefore has more need
  of the processor.
• Furthermore, instead of a queue, CFS uses a
  Red-Black tree to store, sort, and schedule tasks.

33
Priorities and more

• While CFS does not directly use priorities or
  priority queues, it does use them to modulate
  vruntime buildup.
• In this version, priority is inverse to its
  effect a higher priority task will accumulate
  vruntime more slowly, since it needs more CPU
  time.
• Likewise, a low-priority task will have its
  vruntime increase more quickly, causing it to be
  preempted earlier.
• Nice value lower value means higher priority.
• Relative priority, not absolute...

34
RB Trees

• A red-black tree is a binary search tree, which
  means that for each node, the left subtree only
  contains keys less than the node's key, and the
  right subtree contains keys greater than or equal
  to it.
• A red-black tree has further restrictions which
  guarantee that the longest root-leaf path is at
  most twice as long as the shortest root-leaf
  path. This bound on the height makes RB Trees
  more efficient than normal BSTs.
• Operations are in O(log n) time.

35
The CFS Tree

• The key for each node is the vruntime of the
  corresponding task.
• To pick the next task to run, simply take the
  leftmost node.

http//www.ibm.com/developerworks/linux/library/l-
completely-fair-scheduler/
36
Modular scheduling

• Alongside the initial CFS release came the notion
  of modular scheduling, and scheduling classes.
  This allows various scheduling policies to be
  implemented, independent of the generic
  scheduler.
• sched.c, which we have seen, contains that
  generic code. When schedule() is called, it will
  call pick_next_task(), which will look at the
  task's class and call the class-appropriate
  method.
• Let's look at the sched_class struct...(sched.h
  L976)

37
Scheduling classes!

• Two scheduling classes are currently implemented
  sched_fair, and sched_rt.
• sched_fair is CFS, which I've been talking about
  this whole time.
• sched_rt handles real-time processes, and does
  not use CFS it's basically the same as the
  previous scheduler.
• CFS is mainly used for non-real-time tasks.

38
A visual aid is in order...

• Classes are connected via linked-list, making it
  easy to iterate among them.
• Each has its own functions corresponding to the
  core sched_class.

http//www.ibm.com/developerworks/linux/library/l-
completely-fair-scheduler/