Uniprocessor Scheduling

1
Uniprocessor Scheduling

• Chapter 9

2
Levels of Scheduling

• Long-term (high level) A new job is added to the
  pool of available processes.
• Primarily required for batch jobs, which may be
  held on disk in a queue.
• Medium-term (intermediate level) One or more
  processes is swapped in
• Short-term (low-level) decides in what order
  runnable processes (or threads) will be assigned
  to the CPU.

3
Scheduling Levels State Transitions
4
Short-Term Scheduling

• The focus in this chapter is short-term
  scheduling, or processor scheduling.
• Events that invoke the scheduler
• interrupts clock, I/O
• system calls
• signals
• In short, any event that could cause suspension
  or preemption of the current process.

5
CPU-I/O Bursts

• In a uniprogramming system, a process will
  alternate between CPU bursts I/O bursts.
• CPU burst use the processor (execute)
• I/O bursts wait for completion of some event
• Process run time ? (all CPU bursts)
• Process wait time ? (all I/O bursts)
• A process is I/O bound if it has many short CPU
  bursts, and CPU bound if it has fewer, but
  longer, CPU bursts.
• Multiprogramming was devised to take advantage of
  I/O bursts.

6
Preemptive Scheduling

• Scheduling is non-preemptive if new processes are
  dispatched only when a process leaves the Run
  state voluntarily
• process terminates, or moves to Blocked state
• In preemptive scheduling, processes can be
  preempted, or removed from the Run state before
  the end of the current CPU burst.
• Preemption improves response time and
  predictability in an interactive system.

7
Preemptive Scheduling

• Events which may lead to preemption
• a new process (or thread) is created
• an interrupt enables a process to move from the
  blocked to the ready state
• a clock interrupt
• Preemption requires hardware assistance in the
  form of a timer which can be set to interrupt
  periodically.
• Active multiprogramming uses preemption
• Passive multiprogramming non-preemptive

8
Disadvantages

• If a process is preempted while updating shared
  resources or kernel data structures, the data may
  become inconsistent if no synchronization is
  provided.
• Preemptive algorithms increase the number of
  process switches, which adds to system overhead.

9
Short-term Scheduling Criteria

• User oriented criteria
• response time (for interactive processes)
  measured from the time a request is made until
  the time results begin to appear
• turnaround time (for batch processes) the total
  time in the system - time spent waiting plus
  actual execution time
• deadlines
• predictability system load shouldnt affect cost
  or response time.

10
Short-term Scheduling Criteria

• System oriented criteria include
• throughput the number of processes completed per
  unit of time. Sometimes stated as the amount of
  work performed per unit of time.
• processor utilization percentage of time the
  processor is busy (with useful work).
• System-oriented criteria arent very important in
  single user systems. Response time is the key
  factor.

11
Other System Criteria

• Priorities
• intrinsic process priority
• task deadlines
• resource requirements, process execution
  characteristics (e.g., I/O bound processes,
  amount of CPU time received recently)
• Fairness In the absence of priorities, treat all
  processes about the same. When there are
  priorities, treat all processes of the same
  priority about the same.
• Ensure that all processes make progress.
• Fairness in most environments means no
  starvation.
• Balance use of system resources

12
Characterization of Scheduling Policies

• The selection function determines which ready
  process gets to run next
• The decision mode is non-preemptive or preemptive
• Other notation
• w total time in system, waiting executing
• e time spent in execution so far
• s total service time required by process,
  including e
• Tq turnaround time total time process spends in
  system (waiting plus executing)
• Tq/Ts normalized turnaround time (Ts is service
  time)

13
Sample Data Set for Examples
Service Time
Arrival Time
Process
A/P1
0
3
B/P2
2
6
C/P3
4
4
D/P4
6
5
E/P5
8
2
Service time total CPU time needed, or length of
next CPU burst Long jobs have a high service time
(long is relative)
14
First Come First Served (FCFS)

• Selection function the oldest process in the
  ready queue (maxw). Hence, FCFS.
• Decision mode nonpreemptive
• a process runs until it blocks itself
• while it waits, another process can run

15
Characteristics of FCFS

• Wide variation in wait times, sensitive to
  process arrival order.
• Favors long (CPU-bound) processes over short (I/O
  bound)
• Consequently, not effective in an interactive
  environment - normalized turnaround time for
  short jobs can be terrible.
• But easy to implement, no danger of starvation

16
Round-Robin (RR)

• Selection function same as FCFS
• Decision mode preemptive
• a process runs until it blocks or until its time
  slice - typically from 10 to 100 ms - has expired
• a timer is set to interrupt at the end of the
  time slice the running process is put at the end
  of the ready queue
• FCFS and RR are implemented with a FIFO queue

17
Characteristics of Round Robin

• RR is designed to give better service to short
  processes.
• Its appropriate for an interactive or
  transaction processing environments
• Less variance in wait times than FCFS
• Biased against I/O bound jobs, since they may not
  use the entire time slice before blocking, and
  thus get less total CPU time.

18
RR Performance

• The main issue in RR performance is choice of
  quantum size. (quantum time slice)
• Should be significantly larger than process
  switch time (for efficiency)
• Should be slightly longer than the typical
  interaction (to move processes through the queue
  quickly)
• If the quantum is too long, performance
  approaches that of FCFS

19
Time Quantum for Round Robin
In the figures above, we see the difference in
completion time for a process when the quantum is
a) slightly larger than the interaction time and
b) slightly smaller than that time.
20
Virtual Round Robin

• Designed to avoid the bias against I/O bound jobs
  that is found in basic RR.
• Suppose a process blocks for I/O after executing
  for p time units (quantum q units)
• When the process is released from the I/O block
  it is put onto an auxiliary queue instead of the
  normal ready queue.
• Processes are dispatched first from the auxiliary
  queue, with a quantum q? q - p

21
Queuing for Virtual Round Robin (VRR)
22
Shortest Process Next (SPN)

• Selection function the process with the shortest
  expected CPU burst time. (Mins)
• Decision mode nonpreemptive
• Requires estimated CPU burst times
• But note that a short process may still get stuck
  behind a long process since theres no preemption

23
SPN Characteristics

• Optimal for nonpreemptive algorithms
• Minimizes average wait time, maximizes throughput
• Long jobs may be starved
• Difficult to estimate burst times.
• Lack of preemption is not suitable in a
  time-sharing environment-a long job can still
  monopolize the CPU once it is dispatched.

24
Shortest Remaining Time (SRT)

• A preemptive version of SPN
• The scheduler will preempt the current process
  when a shorter job arrives in the ready state.
• New jobs
• Jobs returning from blocked state with reduced
  service time
• Still depends on having time estimates and
  records of elapsed service times (i.e., extra
  overhead)

25
Analysis of SRT

• Somewhat more overhead than SPN
• Better throughput
• Favors short jobs even more than SPN does

26
Highest Response Ratio Next (HRRN)

• Decision mode non-preemptive
• Selection criterion the largest Response Ratio
  RR (w s) / s, where
• s expected service time
• w time spent waiting for processor (notice that
  this is not the same w mentioned earlier, which
  includes service time to date).
• Goal minimize avg. normalized turnaround time
• Avoids starvation while favoring short or old
  jobs.
• Still requires estimated service times.

27
Highest Response Ratio Next (HRRN)

• Choose next process with the greatest ratio

time spent waiting expected service
time expected service time
28
Sample Data Set for Examples
Service Time
Arrival Time
Process
A
0
3
B
2
6
C
4
4
D
6
5
E
8
2
B completes at time 9. C arrives at t4, w5
(ws)/s 9/4 D arrives at t6, w 3 RR
8/5E arrives at t8, RR 3/2. Schedule process
C. C completes at t13 RR for D is now
(75)/512/5 RR for E is (52)/2 7/2. Schedule
process E next.
29
Other Scheduling Algorithms

• Deadline scheduling schedule the job with the
  closest deadline (may be used in real time
  environments)
• Fair-share scheduling process groups get a
  percentage of total CPU time, an individual
  process gets a portion of its groups time.
• Priority queues a prioritized set of FCFS
  queues. Schedule the first process on the
  highest priority non-empty queue.

30
Scheduling Considerations

• In the absence of any knowledge about execution
  time, deadlines, or priorities, how can an
  operating system schedule processes fairly?
• One approach is to use knowledge about time spent
  executing so far.
• A process that has accumulated significant recent
  CPU time could be penalized, on the basis that it
  must be a long job.

31
Multilevel Feedback Scheduling (MFS)

• Preemptive scheduling with dynamic priorities.
• Several FIFO ready queues with decreasing
  priorities Pr(RQ0) gt Pr(RQ1) gt ... gt Pr(RQn)
• New processes are placed at the tail end of RQ0
• Dispatch from the head of RQ0 when quantum
  expires, preempt it and place at the end of RQ1
• In general, if a process is preempted after
  having been on RQi-1, it moves to RQi
• Processes in RQn (lowest queue) execute round
  robin
• Dispatcher chooses a process for execution from
  the highest priority non-empty queue.

32
Feedback Queues

• A running process either completes or moves to a
  lower queue - processes that block may be
  considered new when they return to the Ready
  state, in which case they return to the highest
  level queue.

33
MFS Characteristics

• New processes are favored over old processes
• Short (I/O-bound) processes will complete
  relatively quickly, since they stay in higher
  priority queues
• Long (CPU-bound) jobs will drift downward.
• Starvation is likely for long jobs
• give longer quanta to lower level queues
• age the process allow it to move back up to a
  higher priority queue

34
Time Quantum for Multilevel Feedback Scheduling

• Two examples one with constant quantum value,
  one with increasing quantum for each queue.
• Ex in second figure, time quantum of RQi 2i-1
• It may still be necessary to use aging to
  guarantee timely completion for long processes.

35
(No Transcript)
36
A Traditional Unix Scheduler

• Targeted toward multi-user, time-shared
  environments
• Goal good response time for interactive users,
  but no starvation for long background jobs.
• Scheduling algorithm Multilevel Feedback
• round robin within each priority queue
• 1-second preemption
• priority based on process type and execution
  history
• priority decreases proportional to recent
  processor utilization, increases as the process
  sits in a queue.

37
Traditional UNIX Scheduling

• User processes start at priority 0 and go up
  kernel level processes have negative priorities -
  the smallest negative number is the highest
  priority.
• The amount of CPU time a process has received
  since the last priority computation is used to
  lower the priority of processes that have been
  running.
• Older CPU usage figures decay to raise the
  priority of processes that have been idle.

38
UNIX Priority Computation
Priority (base priority) (recent CPU usage/2)
(nice factor) Pj(i) Basej CPUj(i-1)/2
nicej CPUj(i) CPUj(i-1)/2 , where Pj(i) is
priority of process j at start of interval
i Basej is base priority of process j CPUj(i) is
weighted average CPU utilization by process j
through interval i. Process execution time (if
any) is added into the figure nicej is
user-controllable adjustment factor
39
Summary

• Traditional UNIX scheduling had a built-in ageing
  factor as a process waits, its CPU usage
  figures decay. At every time interval, the
  figure is reduced by one-half.
• Eventually, if a process doesnt execute,
  CPUj(i-1)/2 will be effectively 0, and the
  processs priority will thus increase.

40
(No Transcript)