CPU Scheduling

1
CPU Scheduling

• Vivek Pai / Kai Li
• Princeton University

2
Gedankentexperimenten(Thought Experiments)

• Can you implement semaphores using Mesa-style
  monitors? If so, how?
• Java has monitors, but no condition variables.
  There are equivalents to wait( ) and signal( ),
  but they operate on a per-monitor basis. How does
  this change life?

3
Reading Assignments

• Sections 2.4 and 3.3 in text
• Obviously not due for next Tuesday
• What about the last readings?
• Wanted to put thread issue in context
• Preparation for project
• Am I going to ask annoying details about the man
  pages? No
• Are you responsible for the general concepts? Yes

4
Center For Teaching Learning

• They perform diagnostic of class
• Why?
• Helps me figure out whats going well
• Helps me figure out what can be improved
• Gets you thinking about what works well in your
  own personal situation
• In-class, probably at end of lecture?

5
Implement Semaphores with Mesa-Monitors

• P( s )
• Acquire( s.mutex )
• --s.value
• if (s.value lt 0 )
• Wait(s.mutex,s.cond )
• Release( s.mutex)

• V( s )
• Acquire( s.mutex )
• s.value
• if (s.value gt 0 )
• Signal( s.cond )
• Release( s.mutex)

6
Monitors and Java-like Approach

• P( s )
• Acquire( s.mutex )
• --s.value
• if (s.value lt 0 )
• Wait( )
• Release( s.mutex)

• Code on left not correct
• Small tweak makes it work
• Whats the tweak?
• What impact does it have?

7
CPU Scheduling

• What is it?
• Mechanism policy
• Current mechanism fairly simple
• Policy choices harder
• Is mechanism always simple?
• No, still open for research

8
What Issues are in Policy?

• Fairness
• Flexibility
• High utilization (efficiency)
• Good response time
• Other issues?
• Smooth media playback, device handling

9
What Issues are in Policy?

• Fairness
• All users should get access to CPU
• Amount of CPU should be roughly even
• Somewhat related forward progress
• More of a correctness issue
• Global forward progress needs some fairness

10
What Issues are in Policy?

• High utilization (efficiency)
• Lots of processes (decisions in scheduler?)
• Lots of resources (want full parallelism)
• Different resource profiles
• Issue?
• How do you get the most useful work out of the
  system?

11
What Issues are in Policy?

• Flexibility - variability in job types
• Long versus short
• Interactive versus non-interactive
• I/O-bound versus compute-bound
• Issues?
• Short jobs shouldnt suffer
• Users shouldnt be annoyed

12
Goals and Assumptions

• Goals
• Short response time type on a keyboard
• Maximize throughput operations (jobs) per second
• Minimize overhead of context switches
• Efficient utilization (CPU, memory, disk etc)
• Minimize turnaround batch users wait
• Fairness
• Is there a difference between fairness and avg
  response time?
• What about average response time and throughput?
• Assumptions
• One program per user and one thread per program
• Programs are independent

13
Non-Preemptive FIFO (FCFS) Policy

• What does it mean?
• Run to completion (old days)
• Run until blocked or yield
• Advantages
• Simple
• Disadvantage
• Short jobs behind long jobs

14
Round Robin
Ready queue
Current process

• Each runs a time slice or quantum
• How do you choose time slice?
• Overhead vs. throughputs
• Overhead is typically about 1 or less

15
Is Fairness Always Good?

• Assume 10 jobs, each takes 100 seconds
• Assume no other overhead
• Total CPU time? 1000 seconds, always
• Implications?
• Last job always finishes at 1000 seconds
• So whats the point of scheduling?

16
FIFO Example

• Job 1 start 0, end 100
• Job 2 start 100, end 200
• Job 10 start 900, end 1000
• Average wait time? 100 200 / N
• 550 seconds

17
Round Robin Example

• Assume each quantum is 1 second
• Job 0 0, 10, 20, 30, 40,, 990
• Job 1 1, 11, 21, 31,, 991
• Job 2 2, 12, 22, 32,, 992
• Average wait time? 990991/N
• 995 seconds

18
Like, Whoa! Dude!

• Unfair policy was faster!
• Job 10 always ended at the same time
• Round-Robin just hurt jobs 1-9 with no gain

19
So Why Use Round-Robin?

• Imagine 10 jobs
• Jobs 1-9 are 100 seconds
• Job 10 is 10 seconds
• Which policy is better now?

20
Queuing Theory

• An entire discipline to itself
• Mathematically oriented
• Some neat results
• Systems harder to model these days
• Whats used now? Simulation

21
Adding I/O Into the Mix

• Resource utilization example
• A and B each uses 100 CPU
• C loops forever (1ms CPU and 10ms disk)
• Time slice 100ms nearly 5 of disk utilization
  with Round Robin
• Time slice 1ms nearly 90 of disk utilization
  with Round Robin and nearly 100 of CPU
  utilization
• What do we learn from this example?
• Small time slice can improve utilization

22
Virtual Round Robin
Timeout

• To improve fairness for I/O bound processes
• Aux queue is FIFO
• I/O bound processes go to aux queue (instead of
  ready queue) to get scheduled
• Aux queue has preference over ready queue

Dispatch
CPU
Admit
Aux queue
I/O wait
I/O wait
I/O completion
I/O wait
23
STCF vs. SRTCF

• Shortest time to completion first (shortest job
  first)
• Non-preemptive
• Shortest remaining time to completion first
• Preemptive version
• Advantage
• Minimal average response time
• Disadvantage
• Can cause starvation difficult to know the
  future
• Question
• Can you do better than SRTCF in terms of average
  response time?

24
Priority Scheduling

• The method
• Assign each process a priority
• Run the process with highest priority in the
  ready queue first
• Adjust priority dynamically (I/O wait raises the
  priority)
• Advantage
• Fairness
• Disadvantage
• Users can hit keyboard frequently

25
Multiple Queue (Feedback)
Priority 0 1 2 3
Time slices 1 2 4 8

• Jobs start at highest priority queue
• If timeout expires, drop one level
• If timeout doesnt expires, stay or pushup one
  level

26
Lottery Scheduling

• Motivations
• SRTCF does well with avg response time, but
  unfair
• Adjusting priority is a bit ad hoc. For example,
  what rate?
• Lottery method
• Give each job a number of tickets
• Randomly pick a winning tickets
• To approximate SRTCF, short jobs give more
  tickets
• To avoid starvation, give each job at least one
  ticket
• For more details
• C.A. Waldspurger and W.E. Weihl, Lottery
  Scheduling Flexible Proportional-Share Resource
  Management. Proc. of the 1st USENIX Symp. on
  Operating System Design and Implementation. Nov
  1994.

27
Traditional Unix Scheduling (SVR3, 4.3 BSD)

• Multi-Queue Feedback with 1 second preemption
• 1 sec preemption
• Preempt if a process doesnt block or complete
  within 1 second
• Priority is recomputed every second
• Pi base CPUi-1/ 2 nice, where CPUi (Ui
  CPUi-1) / 2
• Base is the base priority of the process
• Ui is process utilization in interval i
• Priorities
• Swapper
• Block I/O device control
• File operations
• Character I/O device control
• User processes

28
Multiprocessor and Cluster
CPU
CPU
CPU


L1
L1
L1
L2
L2
L2
External Network
Memory

• Cluster/Multicomputer
• Distributed memory
• An OS on each box

• Multiprocessor architecture
• L2 cache coherence
• A Single image OS

29
Multiprocessor/Cluster Scheduling

• New design issue process/thread to processor
  assignment
• Gang scheduling (coscheduling)
• Threads of the same process will run together (on
  a multiprocessor)
• Processes of the same application run together
  (on a cluster)
• Dedicated Processor Assignment
• Threads will be running to specific processors to
  completion
• Is this a good idea?
• Yes cache misses in shared multiprocessor can be
  avoided
• No lack of fairness

30
Real-Time Scheduling

• Two types of real-time
• Hard deadline must meet, otherwise can cause
  fatal error
• Soft headline meet most of the time, but not
  mandatory
• Characteristics
• Deterministic upper bound on when to get
  services on an I/O
• Responsive how long does OS delays before ack an
  interrupt
• User control provide users with abilities to
  control and specify
• Reliability reduced level of service instead of
  reboot

31
Deadline Scheduling

• Admission control
• Take a job only if the system can guarantee
  real-time
• Information needed
• Ready time time at which task becomes ready
• Starting deadline time by which a task must
  begin
• Completion deadline time by which a task must
  complete
• Processing time time required to execute the
  task to completion
• Resource requirements
• Priority
• Subtask structure

32
Rate Monotonic Scheduling (Liu 73)

• Assign priorities to tasks based on basis of
  their periods
• Highest priority with the shortest period
• For earliest deadline first, we can meet all
  deadlines if following hold
• where Ci computation time, and Ti period
• For RMS algorithm, the following holds
• The upper bound is conservative. In practice,
  the utilization is close to 90

Ci
?
? 1
Ti
Ci
?
? N(21/n 1)
Ti
33
Unix SVR4 Scheduling

• Two major modifications
• A static priority scheduler with 3 classes, 160
  priorities, and round-robin within each priority
• Insertion of preemption points identify
  preemptive regions of code
• Three classes
• Real time (159-100) can make preemptive points
  in the kernel
• Kernel (99-60) run before any time-shared
  processes but after real time
• Time shared (59-0) intended for user processes
  that are not real time

Bit vector
0
1
0
1
0

Dispatch queue
i
0
1
2
159
34
Windows NT Scheduling

• Classes and priorities
• Real time 16 static priorities
• Variable 16 variable priorities, start at a
  base priority
• If a process has used up its quantum, lower its
  priority
• If a process waits for an I/O event, raise its
  priority
• Priority-driven scheduler
• For real-time class, do round robin within each
  priority
• For variable class, multiple queue feedback
• Multiprocessor scheduling
• For N processors, run N-1 highest priority
  threads on N-1 processors and run remaining
  threads on a single processor
• A thread will wait for processors in its affinity
  set, if there are other threads available (for
  variable priorities)