CS 4284 Systems Capstone Resource Allocation & Scheduling Godmar Back
Resource Allocation and Scheduling
Resource Allocation & Scheduling Resource management is primary OS function Involves resource allocation & scheduling Who gets to use what resource and for how long Example resources: CPU time Disk bandwidth Network bandwidth RAM Disk space Processes are the principals that use resources often on behalf of users
Preemptible vs Nonpreemptible Resources Nonpreemptible resources: Once allocated, can t easily ask for them back must wait until process returns them (or exits) Examples: Locks, Disk Space, Control of terminal Preemptible resources: Can be taken away ( preempted ) and returned without the process noticing it Examples: CPU, Memory
Physical vs Virtual Memory Classification of a resource as preemptible depends on price one is willing to pay to preempt it Can theoretically preempt most resources via copying & indirection Virtual Memory: mechanism to make physical memory preemptible Take away by swapping to disk, return by reading from disk (possibly swapping out others) Not always tolerable resident portions of kernel Pintos kernel stack pages
Space Sharing vs Time Sharing Space Sharing: Allocation ( how much? ) Use if resource can be split (multiple CPUs, memory, etc.) Use if resource is non-preemptible Time Sharing: Scheduling ( how long? ) Use if resource can t be split Use if resource is easily preemptible
CPU vs. Other Resources CPU is not the only resource that needs to be scheduled Overall system performance depends on efficient use of all resources Resource can be in use (busy) or be unused (idle) Duty cycle: portion of time busy Consider I/O device: busy after receiving I/O request if CPU scheduler delays process that will issue I/O request, I/O device is underutilized Ideal: want to keep all devices busy
Per-process perspective Process alternates between CPU bursts & I/O bursts I/O Bound Process CPU Bound Process CPU I/O
Global perspective If these were executed on the same CPU: I/O Bound Process CPU Bound Process Waiting CPU I/O
CPU Scheduling Part I
CPU Scheduling Terminology A job (sometimes called a task, or a job instance) Activity that s scheduled: process or part of a process Arrival time: time when job arrives Start time: time when job actually starts Finish time: time when job is done Completion time (aka Turn-around time) Finish time Arrival time Response time Time when user sees response Arrival time Execution time (aka cost): time a job needs to execute Arrival Time Start Time Finish Time waiting CPU burst I/O waiting CPU Response Time Completion Time
CPU Scheduling Terminology (2) Waiting time = time when job was readyto-run didn t run because CPU scheduler picked another job Blocked time = time when job was blocked while I/O device is in use Completion time Execution time + Waiting time + Blocked time
Static vs Dynamic Scheduling Static All jobs, their arrival & execution times are known in advance, create a schedule, execute it Used in statically configured systems, such as embedded real-time systems Dynamic or Online Scheduling Jobs are not known in advance, scheduler must make online decision whenever jobs arrives or leaves Execution time may or may not be known Behavior can be modeled by making assumptions about nature of arrival process
Scheduling Algorithms vs Scheduler Implementations Scheduling algorithms properties are (usually) analyzed under static assumptions first; then adapted for dynamic scenarios Algorithms often consider only an abstract notion of (CPU) jobs, but a dynamic scheduler must map that to processes with alternating - and repeating - CPU and IO bursts Often applies static algorithm to current ready queue Algorithms often assume length of job/cpu burst is known, but real scheduler must estimate expected execution cost (or make assumptions)
Preemptive vs Nonpreemptive Q.: when is scheduler asked to pick a thread from ready queue? Nonpreemptive: Only when RUNNING BLOCKED transition Or RUNNING EXIT Or voluntary yield: RUNNING READY Preemptive Also when BLOCKED READY transition Also on timer (forced call to yield upon intr exit) Scheduling Process must wait for event BLOCKED RUNNING Event arrived Process preempted READY Scheduler picks process
CPU Scheduling Goals Minimize latency Can mean (avg) completion time Can mean (avg) response time Maximize throughput Throughput: number of finished jobs per time-unit Implies minimizing overhead (for context-switching, for scheduling algorithm itself) Requires efficient use of non-cpu resources Fairness Minimize variance in waiting time/completion time
Scheduling Constraints Reaching those goals is difficult, because Goals are conflicting: Latency vs. throughput Fairness vs. low overhead Scheduler must operate with incomplete knowledge Execution time may not be known I/O device use may not be known Scheduler must make decision fast Approximate best solution from huge solution space
First Come First Serve Schedule processes in the order in which they arrive Run until completion (or until they block) Simple! Example: Q.: what is the average completion time? 0 2 7 20 22 27
FCFS (cont d) Disadvantage: completion time depends on arrival order Unfair to short jobs Possible Convoy Effect: 1 CPU bound (long CPU bursts, infrequent I/O bursts), multiple I/O bound jobs (frequent I/O bursts, short CPU bursts). CPU bound process monopolizes CPU: I/O devices are idle New I/O requests by I/O bound jobs are only issued when CPU bound job blocks CPU bound job leads convoy of I/O bound processes FCFS not usually used for CPU scheduling, but often used for other resources (network device)
Round-Robin Run process for a timeslice (quantum), then move on to next process, repeat Decreases avg completion if jobs are of different lengths No more unfairness to short jobs! Q.: what is the average completion time? 0 5 8 27
Round Robin (2) What if there are no short jobs? 0 7 14 21 Q.: what is the average completion time? What would it be under FCFS?
Round Robin Cost of Time Slicing Context switching incurs a cost Direct cost (execute scheduler & context switch) + indirect cost (cache & TLB misses) Long time slices lower overhead, but approaches FCFS if processes finish before timeslice expires Short time slices lots of context switches, high overhead Typical cost: context switch < 10µs Time slice typically around 100ms Note: time slice length!= interval between timer interrupts where periodic timers are used
Shortest Process Next (SPN) Idea: remove unfairness towards short processes by always picking the shortest job If done nonpreemptively also known as: Shortest Job First (SJF), Shortest Time to Completion First (STCF) If done preemptively known as: Shortest Remaining Time (SRT), Shortest Remaining Time to Completion First (SRTCF)
SPN (cont d) Provably optimal with respect to avg waiting time: Moving shorter job up reduces its waiting time more than it delays waiting time of longer job that follows Advantage: Good I/O utilization Disadvantage: Can starve long jobs 0 2 7 27 Big Q: How do we know the length of a job?
Practical SPN Usually don t know (remaining) execution time Exception: profiled code in real-time system; or worstcase execution time analysis (WCET) Idea: determine future from past: Assume next CPU burst will be as long as previous CPU burst Or: weigh history using (potentially exponential) average: more recent burst lengths more predictive than past CPU bursts Note: for some resources, we know or can compute length of next job : Example: disk scheduling (shortest-seek time first)
Aside http://cacm.acm.org/magazines/2014/2/17 1680-computation-takes-time-but-howmuch/fulltext Computation Takes Time, But How Much? Reinhard Wilhelm, Daniel Grund Communications of the ACM, Vol. 57 No. 2, Pages 94-103 10.1145/2500886
Multi-Level Feedback Queue Scheduling Kleinrock 1969 Want: preference for short jobs (tends to lead to good I/O utilization) longer timeslices for CPU bound jobs (reduces context-switching overhead) Problem: Don t know type of each process algorithm needs to figure out Use multiple queues queue determines priority usually combined with static priorities (nice values) many variations of this idea exist
Longer Timeslices Higher Priority MAX 4 3 2 MLFQS Processes start in highest queue Process that use up their time slice move down MIN 1 Higher priority queues are served before lower-priority ones - within highest-priority queue, round-robin Only ready processes are in this queue - blocked processes leave queue and reenter same queue on unblock Processes that starve move up
Basic Scheduling: Summary FCFS: simple unfair to short jobs & poor I/O performance (convoy effect) RR: helps short jobs loses when jobs are equal length SPN: optimal average waiting time which, if ignoring blocking time, leads to optimal average completion time unfair to long jobs requires knowing (or guessing) the future MLFQS: approximates SPN without knowing execution time Can still be unfair to long jobs
CPU Scheduling Part II
Case Study: 2.6 Linux Scheduler Variant of MLFQS 140 priorities 0-99 realtime 100-140 nonrealtime Dynamic priority computed from static priority (nice) plus interactivity bonus (pre 2.6.23) Processes scheduled based on dynamic priority SCHED_OTHER Realtime processes scheduled based on static priority SCHED_FIFO SCHED_RR 140 120 100 0 nice=19 nice=0 nice=-20
Linux Scheduler (2) Instead of recomputation loop, recompute priority at end of each timeslice dyn_prio = nice + interactivity bonus (-5 5) Interactivity bonus depends on sleep_avg measures time a process was blocked 2 priority arrays ( active & expired ) in each runqueue (Linux calls ready queues runqueue )
Linux Scheduler (3) struct prio_array { unsigned int nr_active; unsigned long bitmap[bitmap_size]; struct list_head queue[max_prio]; }; typedef struct prio_array prio_array_t; /* find the highest-priority ready thread */ idx = sched_find_first_bit(array->bitmap); queue = array->queue + idx; next = list_entry(queue->next, task_t, run_list); /* Per CPU runqueue */ struct runqueue { prio_array_t *active; prio_array_t *expired; prio_array_t arrays[2]; } Finds highest-priority ready thread quickly Switching active & expired arrays at end of epoch is simple pointer swap ( O(1) claim)
Linux Timeslice Computation Linux scales static priority to timeslice Nice [ -20 0 19 ] maps to [800ms 100 ms 5ms] Various tweaks: interactive processes are reinserted into active array even after timeslice expires Unless processes in expired array are starving processes with long timeslices are roundrobin d with other of equal priority at subtimeslice granularity
History Variants of MLFQS dominant until a few years ago; still used in Windows kernel Accompanied by belief that online scheduler must be O(1) with small c MLFQS are easily manipulated and do not guarantee fair ( proportional ) CPU assignments Another problem is accuracy of accounting sampling charges entire tick to process that happened to be running at that point
Accuracy of accounting Instead of relying on sampling, modern versions of OS use cycle counters or highprecision timers to accurate determine a process s recent CPU usage. This also makes it easy to exclude time spent in IRQ handling that should not be charged to a process See [Inside the Vista Kernel] for a graph
Linux s CFS Linux went a step further and reinvented WFQ (of course without any credit), implemented in its CFS completely fair scheduler O (log n) red/black tree But does not aim to support really large n But, as we ll see, WFQ does not automatically give precedence to I/O bound apps; required a lot of Interactivity improvements to tune it heuristically Well-known trade-off between fairness & latency
Proportional Share Scheduling Aka Fair-Share Scheduling None of algorithms discussed so far provide a direct way of assigning CPU shares E.g., give 30% of CPU to process A, 70% to process B Proportional Share algorithms do by assigning tickets or shares to processes Process get to use resource in proportion of their shares to total number of shares Lottery Scheduling, Weighted Fair Queuing/Stride Scheduling [Waldspurger 1995]
Lottery Scheduling Idea: number tickets between 1 N every process gets p i tickets according to importance process 1 gets tickets [1 p 1-1] process 2 gets tickets [p 1 p 1+ p 2-1] and so on. Scheduling decision: Hold a lottery and draw ticket, holder gets to run for next time slice Nondeterministic algorithm Q.: how to implement priority donation?
Weighted Fair Queuing Uses per process virtual time Increments process s virtual time by a stride after each quantum, which is defined as (process_share) -1 Choose process with lowest virtual finishing time virtual finishing time is virtual time + stride Also known as stride scheduling Linux now implements a variant of WFQ/Stride Scheduling as its CFS completely fair scheduler
WFQ Example (A=3, B=2, C=1) Ready Queue is sorted by Virtual Finish Time (Virtual Time at end of quantum if a process were scheduled) Time Task A Task B Task C Ready Queue Who Runs 0 1/3 1/2 1 A (1/3) B (1/2) C (1) A 1 2/3 1/2 1 B (1/2) A (2/3) C (1) B 2 2/3 1 1 A (2/3) C(1) B(1) A 3 1 1 1 C(1) B(1) A(1) C 4 1 1 2 B(1) A(1) C(2) B 5 1 3/2 2 A(1) B(3/2) C(2) A 6 4/3 3/2 2 A (4/3) B(3/2) C(2) One scheduling epoch. A ran 3 out of 6 quanta, B 2 out of 6, C 1 out of 6. This process will repeat, yielding proportional fairness.
WFQ (cont d) WFQ requires a sorted ready queue Linux now uses R/B tree Higher complexity than O(1) linked lists, but appears manageable for real-world ready queue sizes Unblocked processes that reenter the ready queue are assigned a virtual time reflecting the value that their virtual time counter would have if they d received CPU time proportionally Accommodating I/O bound processes still requires fudging In strict WFQ, only way to improve latency is to set number of shares high but this is disastrous if process is not truly I/O bound Linux uses sleeper fairness, to identify when to boost virtual time; similar to the sleep average in old scheduler
Linux SMP Load Balancing Runqueue is per CPU Periodically, lengths of runqueues on different CPU is compared Processes are migrated to balance load Aside: Migrating requires locks on both runqueues static void double_rq_lock( runqueue_t *rq1, runqueue_t *rq2) { if (rq1 == rq2) { spin_lock(&rq1->lock); } else { if (rq1 < rq2) { spin_lock(&rq1->lock); spin_lock(&rq2->lock); } else { spin_lock(&rq2->lock); spin_lock(&rq1->lock); } } }
Real-time Scheduling Real-time systems must observe not only execution time, but a deadline as well Jobs must finish by deadline But turn-around time is usually less important Common scenario are recurring jobs E.g., need 3 ms every 10 ms (here, 10ms is the recurrence period T, 3 ms is the cost C) Possible strategies RMA (Rate Monotonic) Map periods to priorities, fixed, static EDF (Earliest Deadline First) Always run what s due next, dynamic
EDF Example Task T C A 4 1 B 8 4 C 12 3 Assume deadline equals period (T). A B C 0 5 10 15 20 25 Hyper-period
EDF Example Task T C A 4 1 B 8 4 C 12 3 Assume deadline equals period (T). A B C 0 5 10 15 20 25
EDF Example Task T C A 4 1 B 8 4 C 12 3 Assume deadline equals period (T). A B C 0 5 10 15 20 25
EDF Example Lexical order tie breaker (C > B > A) Task T C A 4 1 B 8 4 C 12 3 Assume deadline equals period (T). A B C 0 5 10 15 20 25
EDF Example Task T C A 4 1 B 8 4 C 12 3 Assume deadline equals period (T). A B C 0 5 10 15 20 25
EDF Example Task T C A 4 1 B 8 4 C 12 3 Assume deadline equals period (T). A B C 0 5 10 15 20 25
EDF Example Task T C A 4 1 B 8 4 C 12 3 Assume deadline equals period (T). A B C 0 5 10 15 20 25
EDF Example Task T C A 4 1 B 8 4 C 12 3 Assume deadline equals period (T). A B C 0 5 10 15 20 25
EDF Example Task T C A 4 1 B 8 4 C 12 3 Assume deadline equals period (T). A B C 0 5 10 15 20 25 Pattern repeats
EDF Properties Feasibility test: U = 100% in example Bound theoretical Sufficient and necessary Optimal
Scheduling Summary OS must schedule all resources in a system CPU, Disk, Network, etc. CPU Scheduling affects indirectly scheduling of other devices Goals for general purpose schedulers: Minimizing latency (avg. completion or waiting time) Maximing throughput Provide fairness In Practice: some theory, lots of tweaking