LEC5 CPU Scheduling - ranger.uta.eduranger.uta.edu/~jrao/CSE3320/spring2018/slides/LEC5_CPU_Scheduling.pdfMultilevel Feedback Queue. Multilevel Feedback Queue 16 1/31/17 RR, q=4 RR,

1/31/171

CSE3320OperatingSystemsCPUScheduling

Jia RaoDepartmentofComputerScience and Engineering

http://ranger.uta.edu/~jrao

WhatisCPUScheduling?

1/31/172

• Thefive-stateprocessmodel

ready running

waiting

new terminated

schedulerdispatch

interrupt

I/OoreventwaitI/Ooreventcompletion

admit exit

CPUschedulingSelectsfromamongtheprocesses/threadsthatarereadytoexecute,andallocatestheCPUtoit

WhyCPUScheduling?

1/31/173

• Insupportofmultiprogrammingo uniprocessorsystems

} Time-sharingprocessor

o multiprocessorsystems} Efficientlydistributingtasks

o Real-timesystems} Reliablyguaranteeingdeadlines

• Itis(maybe)themostimportantpartinaOSo WhysomeOSseemstobefasterthanothers?o WhyIdonotseeperformanceimprovementwhenupgradingtoa16-corecomputer?

InthisLecture

1/31/174

• Outlineo BasicsofCPUscheduling

} Schedulingpolicies} Evaluationcriteria} Examples} Apracticalpolicy

o Challengesonemerginghardwareandapplications} Manycore,NUMA,asymmetricprocessors} Datacenter,accurateresourceprovisioning

o Acloselookatthestate-of-art} TheLinuxCFSscheduler

CPUScheduling

1/31/175

• CPUschedulingmaytakeplaceato Clockinterruptso I/Ocompletiono I/Ointerruptso Termination

• Nonpreemptiveo Schedulingonlywhencurrentprocessterminatesorgivesupcontrol

• Preemptiveo Processescanbeforcedtogiveupcontrol

processorpick_next_task()readyqueue

preemption

exit/blocked

preemptive

nonpreemptive

SchedulingGoals

1/31/176

SchedulingGoals:A DifferentPointofView

1/31/177

• Userorientedàminimizeo Responsetime(waittime):thetimethatthefirstresponseisreceived(interactivity)

o Turnaroundtime:thetimethatthetaskfinisheso Predictability:variationsindifferentruns

• Systemorientedàmaximizeo Throughput:#oftasksthatfinishpertimeunito Utilization:thepercentageoftimetheCPUisbusyo Fairness:avoidstarvation

ProcessBehaviors

1/31/178

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

Ø I/O is when a process enters the blocked state waiting for an external device to complete its work

SchedulingPolicies

1/31/179

• BatchSystemso First-ComeFirst-Serveo ShortestJobFirsto ShortestRemainingTimeNext

• InteractiveSystemso Round-Robino PrioritySchedulingo MultipleQueueso ShortestProcessNexto GuaranteedSchedulingo LotteryScheduling

• Real-timeSystemso RateMonotonicSchedulingo EarliestDeadlineFirstScheduling

MoreonSchedulingPolicy

1/31/17*TheactualmajorschedulefunctioninLinux10

• Determinethenextreadytasktoruno Howwedesignpick_next_task()

• Basicpolicieso First-Come,First-Served(FCFS)o Shortest-Job-First(SJF)o RoundRobin(RR)o Priorityscheduling

processorpick_next_task()*readyqueue

preemption

exit/blocked

First-Come,First-Serve(FCFS)

1/31/1711

• CPUschedulesthetaskthatarrivedearliest,non-preemptiveProcess ArrivalTime BurstTime

P1 0 8P2 1 4P3 2 9P4 3 5

260 8 12 21P1 P2 P3 P4

Averageturnaroundtime=((8-0)+(12-1)+(21-2)+(26-3))/4=15.25Averageresponsetime=(0+(8-1)+(12-2)+(21-3))/4=8.75

260 5 14 18P1P2P3P4


ShortestJobFirst(SJF)

1/31/1712

• CPUschedulesthetaskwiththeshortestremainingtimeProcess ArrivalTime BurstTime

P1 0 8P2 1 4P3 2 9P4 3 5


260 8 12 17P1 P2 P3P4nonpreemptive

260 5 10 17Averageturnaroundtime=((17-0)+(5-1)+(10-3)+(26-2))/4=13Averageresponsetime=(0+(1-1)+(5-3)+(17-2))/4=4.25

preemptive P1P2 P3P4P11

RoundRobin(RR)

1/31/1713

• LikeFCFS,butwithlimitedtimeslices,preemptiveProcess ArrivalTime BurstTime

P1 0 8P2 1 4P3 2 9P4 3 5

250 12 16 24Averageturnaroundtime=((20-0)+(8-1)+(26-2)+(25-3))/4=18.25Averageresponsetime=(0+(4-1)+(8-2)+(12-3))/4=4.5

q=4 P1 P2 P3 P4 P1 P3 P3P44 8 20 26

260 9 14 19Averageturnaroundtime=((22-0)+(9-1)+(26-2)+(19-3))/4=17.5Averageresponsetime=(0+(5-1)+(9-2)+(14-3))/4=5.5

q=55

P4P1 P2 P3 P1 P322

PriorityScheduling

1/31/1714

• CPUschedulesthehighestpriorityfirst,FCFSwithinthesamepriority Process Priority BurstTime

P1 3 8P2 1 4P3 4 9P4 2 5

P1P2 P3P4Process Priority BurstTime

P1 2 8P2 4 4P3 1 9P4 3 5

P1 P2P3 P4

Putittogether

1/31/1715

Turnaroundtime Responsetime

FCFS 15.25 8.75

SJF-preemptive 13 4.25

RR(q=5) 17.5 5.5

Priorityscheduling N/A N/A

Throughput Responsetime Starvation

FCFS TBD TBD No

SJF-preemptive High Good Yes

RR Can below Good No

Priorityscheduling Canbehigh Canbegood Can remove

MultilevelFeedbackQueue

MultilevelFeedbackQueue

1/31/1716

RR,q=4

RR,q=8

RR,q=16

Highpriority/shorttask

Lowpriority/longtask

pick_next_task()

WindowsXP,MacOSX,Linux2.6.22andbefore

Anti-StarvationAswaittimeincreases,

ataskgraduallymovesup

Real-timeScheduling

1/31/1717

Schedulable real-time system• Given

o m periodic eventso event i occurs within period Pi and requires Ci seconds

• Then the load can only be handled if

• Example: a soft real-time system with three periodic events, with periods of 100, 200, and 500 ms, respectively. If these events require 50, 30, and 100 ms of CPU time per event, respective, the system is schedulableo Process/context switching overhead is often an issue though!o Given the example, what would be the maximum CPU burst for a 4th

job with a period of 500 ms ?

11

mi

i i

CP=

£å

Misc.

1/31/1718

• I/Otaskso Identify:runsfew,sleepsaloto Consideredasshorttasko Highpriority

• Theoreticalanalysiso Assumetaskdistributiono Queuingmodel

RR,q=4

RR,q=8

RR,q=16

Kerneltasks

I/Otasks

computationtasks

ChallengesonEmergingHardwareandApplications

1/31/1719

• MultiprocessoràMany core


processor

…

Multiprocessor=morepowerfulprocessor

pick_next_task()willbethebottleneck


processor

…readyqueue

…

pick_next_task()

LoadbalancingCacheaffinity

Problem:Self-scheduling,globalcontrolisdifficult

Linux2.4.x

Linux2.6.x

ChallengesonEmergingHardwareandApplications(cont’)

1/31/1720

• NUMA,Asymmetricprocessorso OLD:CPUtimeà usefulworko NEW:calibratedCPUtimeàusefulwork

• Datacenter,accurateresourceprovisioningo Proportionalfairsharingà P1 : P2 =1:2

RR,q=4

RR,q=8

RR,q=16

pick_next_task()

q=?Whentomovetask?Heuristic-based

Fine-grainedAdaptiveq

ACloseLookattheState-of-Art

1/31/1721

• LinuxCompletelyFairScheduler(CFS)o Separatereadyqueueperprocessoro Red-blacktreebasedreadyqueueo Proportionalfairsharing

pick_next_task()Removetheleftmosttask;Runthetask;Everyticksupdatevruntime;vruntime =vruntime +runtime;ifvruntime >curr_leftmost’sPreemptcurr task;Putitbacktothesortedtree;Runcurr_leftmost;

vruntime

Fine-grainAdaptiveq

Priority:lowpriorityTask’svruntime runs

FasterNostarvation

Summary

1/31/1722

• Thebasicschedulingpoliciesareimportanto MultilevelFeedbackQueue=RR+SJF+Priorityo CFS=RR+SJF+Priority+smartdatastructure

• Additionalreadingso Gotohttp://lxr.linux.no/linux+v2.6.24/kernel/o Read/kernel/sched.c,/kernel/sched_fair.c,/include/linux/sched.h(startingfromtheschedule(void)function)

o Seehowthevruntime isactuallyupdatedo Documentation:http://www.kernel.org/doc/Documentation/scheduler/sched-design-CFS.txt

o Anotherinterestingscheduler:BFS} http://ck.kolivas.org/patches/bfs/bfs-faq.txt

Documents

LEC5 CPU Scheduling - ranger.uta.eduranger.uta.edu/~jrao/CSE3320/spring2018/slides/LEC5_CPU_Scheduling.pdfMultilevel Feedback Queue. Multilevel Feedback Queue 16 1/31/17 RR, q=4 RR,