Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
CS162 Operating Systems andSystems Programming
Lecture 10
Scheduling (conāt)
February 25th, 2020Prof. John Kubiatowicz
http://cs162.eecs.Berkeley.edu
Acknowledgments: Lecture slides are from the Operating Systems course taught by John Kubiatowicz at Berkeley, with few minor updates/changes. When slides are obtained from other sources, a a reference will be noted on the bottom of that slide, in which case a full list of references is provided on the last slide.
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 2
Recall: Scheduling
ā¢ Discussion of Scheduling: ā Which thread should run on the CPU next?
ā¢ Scheduling goals, policiesā¢ Look at a number of different schedulers
if ( readyThreads(TCBs) ) {nextTCB = selectThread(TCBs);run( nextTCB );
} else {run_idle_thread();
}
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 3
Recall: Scheduling Policy Goals/Criteriaā¢ Minimize Response Time
ā Minimize elapsed time to do an operation (or job)ā Response time is what the user sees:Ā» Time to echo a keystroke in editorĀ» Time to compile a programĀ» Real-time Tasks: Must meet deadlines imposed by World
ā¢ Maximize Throughputā Maximize operations (or jobs) per secondā Throughput related to response time, but not identical:Ā» Minimizing response time will lead to more context switching than if you
only maximized throughputā Two parts to maximizing throughputĀ» Minimize overhead (for example, context-switching)Ā» Efficient use of resources (CPU, disk, memory, etc)
ā¢ Fairnessā Share CPU among users in some equitable wayā Fairness is not minimizing average response time:Ā» Better average response time by making system less fair
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 4
Recall: Example of RR with Time Quantum = 20ā¢ Example: Process Burst Time
P1 53 P2 8 P3 68 P4 24
ā The Gantt chart is:
ā Waiting time for P1=(68-20)+(112-88)=72P2=(20-0)=20
P3=(28-0)+(88-48)+(125-108)=85P4=(48-0)+(108-68)=88
ā Average waiting time = (72+20+85+88)/4=66Ā¼ā Average completion time = (125+28+153+112)/4 = 104Ā½
ā¢ Thus, Round-Robin Pros and Cons:ā Better for short jobs, Fair (+)ā Context-switching time adds up for long jobs (-)
P10 20
P228
P348
P468
P188
P3108
P4 P1 P3 P3112 125 145 153
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 5
Comparisons between FCFS and Round Robinā¢ Assuming zero-cost context-switching time, is RR always better than
FCFS?ā¢ Simple example: 10 jobs, each take 100s of CPU time
RR scheduler quantum of 1s All jobs start at the same time
ā¢ Completion Times:
ā Both RR and FCFS finish at the same timeā Average response time is much worse under RR!Ā» Bad when all jobs same length
ā¢ Also: Cache state must be shared between all jobs with RR but can be devoted to each job with FIFOā Total time for RR longer even for zero-cost switch!
Job # FIFO RR
1 100 991
2 200 992
ā¦ ā¦ ā¦
9 900 999
10 1000 1000
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 6
Quantum
CompletionTime
WaitTime
AverageP4P3P2P1
Earlier Example with Different Time QuantumP2[8]
P4[24]
P1[53]
P3[68]
0 8 32 85 153
Best FCFS:
6257852284Q = 1
104Ā½11215328125Q = 20
100Ā½8115330137Q = 1
66Ā¼ 88852072Q = 20
31Ā¼885032Best FCFS
121Ā¾14568153121Worst FCFS
69Ā½32153885Best FCFS83Ā½121014568Worst FCFS
95Ā½8015316133Q = 8
57Ā¼5685880Q = 8
99Ā½9215318135Q = 10
99Ā½8215328135Q = 5
61Ā¼68851082Q = 10
61Ā¼58852082Q = 5
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 7
Handling Differences in Importance: Strict Priority Scheduling
ā¢ Execution Planā Always execute highest-priority runable jobs to completionā Each queue can be processed in RR with some time-quantum
ā¢ Problems:ā Starvation:
Ā» Lower priority jobs donāt get to run because higher priority jobsā Deadlock: Priority Inversion
Ā» Not strictly a problem with priority scheduling, but happens when low priority task has lock needed by high-priority task
Ā» Usually involves third, intermediate priority task that keeps running even though high-priority task should be running
ā¢ How to fix problems?ā Dynamic priorities ā adjust base-level priority up or down based on heuristics
about interactivity, locking, burst behavior, etcā¦
Priority 3
Priority 2
Priority 1
Priority 0 Job 5 Job 6
Job 1 Job 2 Job 3
Job 7
Job 4
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 8
Scheduling Fairness
ā¢ What about fairness?ā Strict fixed-priority scheduling between queues is unfair (run
highest, then next, etc):Ā» long running jobs may never get CPU Ā» Urban legend: In Multics, shut down machine, found 10-year-
old job ā Ok, probably notā¦ā Must give long-running jobs a fraction of the CPU even when
there are shorter jobs to runā Tradeoff: fairness gained by hurting avg response time!
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 9
Scheduling Fairness
ā¢ How to implement fairness?ā Could give each queue some fraction of the CPU
Ā» What if one long-running job and 100 short-running ones?Ā» Like express lanes in a supermarketāsometimes express
lanes get so long, get better service by going into one of the other lines
ā Could increase priority of jobs that donāt get serviceĀ» What is done in some variants of UNIXĀ» This is ad hocāwhat rate should you increase priorities?Ā» And, as system gets overloaded, no job gets CPU time, so
everyone increases in priorityāInteractive jobs suffer
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 10
Lottery Scheduling
ā¢ Yet another alternative: Lottery Schedulingā Give each job some number of lottery ticketsā On each time slice, randomly pick a winning ticket
Ā» NOTE: Not a ārealā random number generator; instead pseudo-random number generators can make sure that every ticket picked once before repeating!
ā On average, CPU time is proportional to number of tickets given to each job
ā¢ How to assign tickets?ā To help with responsiveness, give short running jobs more
tickets, long running jobs get fewer ticketsā To avoid starvation, every job gets at least one ticket
(everyone makes progress)ā¢ Advantage over strict priority scheduling: behaves gracefully as
load changesā Adding or deleting a job affects all jobs proportionally,
independent of how many tickets each job possesses
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 11
Lottery Scheduling Example (Cont.)
ā¢ Lottery Scheduling Exampleā Assume short jobs get 10 tickets, long jobs get 1 ticket
ā What if too many short jobs to give reasonable response time?
Ā» If load average is 100, hard to make progressĀ» One approach: log some user out
# short jobs/# long jobs
% of CPU each short jobs gets
% of CPU each long jobs gets
1/1 91% 9%
0/2 N/A 50%
2/0 50% N/A
10/1 9.9% 0.99%
1/10 50% 5%
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 12
How to Evaluate a Scheduling algorithm?ā¢ Deterministic modeling
ā takes a predetermined workload and compute the performance of each algorithm for that workload
ā¢ Queueing modelsā Mathematical approach for handling stochastic workloads
ā¢ Implementation/Simulation:ā Build system which allows actual algorithms to be run against
actual data ā most flexible/general
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 13
How to Handle Simultaneous: Mix of Diff Types of Apps?
ā¢ Consider mix of interactive and high throughput apps:ā How to best schedule them?ā How to recognize one from the other?
Ā» Do you trust app to say that it is āinteractiveā?ā Should you schedule the set of apps identically on servers, workstations, pads,
and cellphones?ā¢ For instance, is Burst Time (observed) useful to decide which application gets
CPU time?ā Short Bursts ā Interactivity ā High Priority?
ā¢ Assumptions encoded into many schedulers:ā Apps that sleep a lot and have short bursts must be interactive apps ā they
should get high priorityā Apps that compute a lot should get low(er?) priority, since they wonāt notice
intermittent bursts from interactive appsā¢ Hard to characterize apps:ā What about apps that sleep for a long time, but then compute for a long time?ā Or, what about apps that must run under all circumstances (say periodically)
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 14
What if we Knew the Future?ā¢ Could we always mirror best FCFS?ā¢ Shortest Job First (SJF):
ā Run whatever job has least amount of computation to do
ā Sometimes called āShortest Time to Completion Firstā (STCF)ā¢ Shortest Remaining Time First (SRTF):
ā Preemptive version of SJF: if job arrives and has a shorter time to completion than the remaining time on the current job, immediately preempt CPU
ā Sometimes called āShortest Remaining Time to Completion Firstā (SRTCF)
ā¢ These can be applied to whole program or current CPU burstā Idea is to get short jobs out of the systemā Big effect on short jobs, only small effect on long onesā Result is better average response time
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 15
Discussion
ā¢ SJF/SRTF are the best you can do at minimizing average response time
ā Provably optimal (SJF among non-preemptive, SRTF among preemptive)
ā Since SRTF is always at least as good as SJF, focus on SRTF
ā¢ Comparison of SRTF with FCFSā What if all jobs the same length?Ā» SRTF becomes the same as FCFS (i.e. FCFS is best can do if all jobs the
same length)ā What if jobs have varying length?Ā» SRTF: short jobs not stuck behind long ones
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 16
Example to illustrate benefits of SRTF
ā¢ Three jobs:ā A, B: both CPU bound, run for week
C: I/O bound, loop 1ms CPU, 9ms disk I/Oā If only one at a time, C uses 90% of the disk, A or B could use 100%
of the CPUā¢ With FCFS:
ā Once A or B get in, keep CPU for two weeksā¢ What about RR or SRTF?
ā Easier to see with a timeline
C
Cās I/O
Cās I/O
Cās I/O
A or B
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 17
SRTF Example continued:
Cās I/O
CABABā¦ C
Cās I/O
RR 1ms time slice
Cās I/O
Cās I/O
CA BC
RR 100ms time slice
Cās I/O
AC
Cās I/O
AA
SRTF
Disk Utilization:~90% but lots of wakeups!
Disk Utilization:90%
Disk Utilization:9/201 ~ 4.5%
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 18
ā¢ Starvationā SRTF can lead to starvation if many small jobs!ā Large jobs never get to run
ā¢ Somehow need to predict futureā How can we do this? ā Some systems ask the user
Ā» When you submit a job, have to say how long it will takeĀ» To stop cheating, system kills job if takes too long
ā But: hard to predict jobās runtime even for non-malicious usersā¢ Bottom line, canāt really know how long job will takeā However, can use SRTF as a yardstick
for measuring other policiesā Optimal, so canāt do any better
ā¢ SRTF Pros & Consā Optimal (average response time) (+)ā Hard to predict future (-)ā Unfair (-)
SRTF Further discussion
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 19
Predicting the Length of the Next CPU Burst
ā¢ Adaptive: Changing policy based on past behaviorā CPU scheduling, in virtual memory, in file systems, etcā Works because programs have predictable behavior
Ā» If program was I/O bound in past, likely in futureĀ» If computer behavior were random, wouldnāt help
ā¢ Example: SRTF with estimated burst lengthā Use an estimator function on previous bursts:
Let tn-1, tn-2, tn-3, etc. be previous CPU burst lengths. Estimate next burst Ļn = f(tn-1, tn-2, tn-3, ā¦)ā Function f could be one of many different time series estimation schemes
(Kalman filters, etc)ā For instance,
exponential averagingĻn = Ī±tn-1+(1-Ī±)Ļn-1 with (0<Ī±ā¤1)
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 20
Multi-Level Feedback Scheduling
ā¢ Another method for exploiting past behavior (first use in CTSS)ā Multiple queues, each with different priority
Ā» Higher priority queues often considered āforegroundā tasksā Each queue has its own scheduling algorithm
Ā» e.g. foreground ā RR, background ā FCFSĀ» Sometimes multiple RR priorities with quantum increasing exponentially (highest:
1ms, next: 2ms, next: 4ms, etc)ā¢ Adjust each jobās priority as follows (details vary)ā Job starts in highest priority queueā If timeout expires, drop one levelā If timeout doesnāt expire, push up one level (or to top)
Long-Running ComputeTasks Demoted to
Low Priority
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 21
Scheduling Details
ā¢ Result approximates SRTF:ā CPU bound jobs drop like a rockā Short-running I/O bound jobs stay near top
ā¢ Scheduling must be done between the queuesā Fixed priority scheduling: Ā» serve all from highest priority, then next priority, etc.
ā Time slice:Ā» each queue gets a certain amount of CPU time Ā» e.g., 70% to highest, 20% next, 10% lowest
Long-Running ComputeTasks Demoted to
Low Priority
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 22
Scheduling Details
ā¢ Countermeasure: user action that can foil intent of the OS designers
ā For multilevel feedback, put in a bunch of meaningless I/O to keep jobās priority high
ā Of course, if everyone did this, wouldnāt work!ā¢ Example of Othello program:
ā Playing against competitor, so key was to do computing at higher priority the competitors. Ā» Put in printf ās, ran much faster!
Long-Running ComputeTasks Demoted to
Low Priority
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 23
Case Study: Linux O(1) Scheduler
ā¢ Priority-based scheduler : 140 prioritiesā 40 for āuser tasksā (set by āniceā), 100 for āRealtime/Kernelāā Lower priority value ā higher priority (for nice values)ā Highest priority value ā Lower priority (for realtime values)ā All algorithms O(1)
Ā» Timeslices/priorities/interactivity credits all computed when job finishes time sliceĀ» 140-bit bit mask indicates presence or absence of job at given priority level
ā¢ Two separate priority queues: āactiveā and āexpiredāā All tasks in the active queue use up their timeslices and get placed on the
expired queue, after which queues swappedā¢ Timeslice depends on priority ā linearly mapped onto timeslice range
ā Like a multi-level queue (one queue per priority) with different timeslice at each levelā Execution split into āTimeslice Granularityā chunks ā round robin through
priority
Kernel/Realtime Tasks User Tasks
0 100 139
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 24
O(1) Scheduler Continuedā¢ Heuristics ā User-task priority adjusted Ā±5 based on heuristics
Ā» p->sleep_avg = sleep_time ā run_timeĀ» Higher sleep_avg ā more I/O bound the task, more reward (and vice versa)
ā Interactive CreditĀ» Earned when a task sleeps for a ālongā timeĀ» Spend when a task runs for a ālongā timeĀ» IC is used to provide hysteresis to avoid changing interactivity for temporary
changes in behaviorā However, āinteractive tasksā get special dispensation
Ā» To try to maintain interactivityĀ» Placed back into active queue, unless some other task has been starved for too
longā¦ā¢ Real-Time Tasksā Always preempt non-RT tasksā No dynamic adjustment of prioritiesā Scheduling schemes:
Ā» SCHED_FIFO: preempts other tasks, no timeslice limitĀ» SCHED_RR: preempts normal tasks, RR scheduling amongst tasks of same
priority
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 25
Linux Completely Fair Scheduler (CFS)ā¢ First appeared in 2.6.23, modified in 2.6.24ā¢ āCFS doesn't track sleeping time and doesn't use heuristics to
identify interactive tasksāit just makes sure every process gets a fair share of CPU within a set amount of time given the number of runnable processes on the CPU.ā
ā¢ Inspired by Networking āFair Queueingāā Each process given their fair share of resourcesā Models an āideal multitasking processorā in which N processes
execute simultaneously as if they truly got 1/N of the processorĀ» Tries to give each process an equal fraction of the processor
ā Priorities reflected by weights such that increasing a taskās priority by 1 always gives the same fractional increase in CPU time ā regardless of current priority
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 26
Real-Time Scheduling (RTS)ā¢ Efficiency is important but predictability is essential:ā We need to predict with confidence worst case response times for systemsā In RTS, performance guarantees are:
Ā» Task- and/or class centric and often ensured a prioriā In conventional systems, performance is:
Ā» System/throughput oriented with post-processing (ā¦ wait and see ā¦)ā Real-time is about enforcing predictability, and does not equal fast computing!!!
ā¢ Hard Real-Timeā Attempt to meet all deadlinesā EDF (Earliest Deadline First), LLF (Least Laxity First),
RMS (Rate-Monotonic Scheduling), DM (Deadline Monotonic Scheduling)ā¢ Soft Real-Time
ā Attempt to meet deadlines with high probabilityā Minimize miss ratio / maximize completion ratio (firm real-time)ā Important for multimedia applicationsā CBS (Constant Bandwidth Server)
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 27
Example: Workload Characteristics
ā¢ Tasks are preemptable, independent with arbitrary arrival (=release) times
ā¢ Tasks have deadlines (D) and known computation times (C) ā¢ Example Setup:
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 28
Example: Round-Robin Scheduling Doesnāt Work
Time
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 29
ā¢ Tasks periodic with period P and computation C in each period: ( , ) for each task
ā¢ Preemptive priority-based dynamic scheduling:ā Each task is assigned a (current) priority based on how close the
absolute deadline is (i.e. for each task!)ā The scheduler always schedules the active task with the closest absolute
deadline
ā¢ Schedulable when
ššš¶š š
š·š”+1š = š·š”
š + šš
š
āš=1 ( š¶š
šš ) ā¤ 1
Earliest Deadline First (EDF)
0 5 10 15
)1,4(1 =T
)2,5(2 =T
)2,7(3 =T
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 30
Choosing the Right Scheduler
I Care About: Then Choose:
CPU Throughput FCFS
Avg. Response Time SRTF Approximation
I/O Throughput SRTF Approximation
Fairness (CPU Time) Linux CFS
Fairness ā Wait Time to Get CPU
Round Robin
Meeting Deadlines EDF
Favoring Important Tasks Priority
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 31
A Final Word On Schedulingā¢ When do the details of the scheduling policy and fairness really
matter?ā When there arenāt enough resources to go around
ā¢ When should you simply buy a faster computer?ā (Or network link, or expanded highway, or ā¦)ā One approach: Buy it when it will pay
for itself in improved response timeĀ» Perhaps youāre paying for worse response
time in reduced productivity, customer angst, etcā¦Ā» Might think that you should buy a faster X
when X is utilized 100%, but usually, response time goes to infinity as utilizationā100%
ā¢ An interesting implication of this curve:ā Most scheduling algorithms work fine in the ālinearā portion of the load
curve, fail otherwiseā Argues for buying a faster X when hit ākneeā of curve
Utilization
Responsetim
e 100%
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 32
Summary (1 of 2)ā¢ Scheduling Goals:
ā Minimize Response Time (e.g. for human interaction)ā Maximize Throughput (e.g. for large computations)ā Fairness (e.g. Proper Sharing of Resources)ā Predictability (e.g. Hard/Soft Realtime)
ā¢ Round-Robin Scheduling: ā Give each thread a small amount of CPU time when it executes; cycle
between all ready threadsā Pros: Better for short jobs
ā¢ Shortest Job First (SJF)/Shortest Remaining Time First (SRTF):ā Run whatever job has the least amount of computation to do/least
remaining amount of computation to doā Pros: Optimal (average response time) ā Cons: Hard to predict future, Unfair
ā¢ Multi-Level Feedback Scheduling:ā Multiple queues of different priorities and scheduling algorithmsā Automatic promotion/demotion of process priority in order to
approximate SJF/SRTF
2/25/2020 Kubiatowicz CS162 Ā©UCB Fall 2020 33
Summary (2 of 2)ā¢ Lottery Scheduling:
ā Give each thread a priority-dependent number of tokens (short tasksāmore tokens)
ā¢ Linux CFS Scheduler : Fair fraction of CPUā Approximates a āidealā multitasking processor
ā¢ Realtime Schedulers such as EDFā Guaranteed behavior by meeting deadlinesā Realtime tasks defined by tuple of compute time and periodā Schedulability test: is it possible to meet deadlines with proposed set of
processes?