CPU Scheduling CPU Scheduling n Basic Concepts n

  • Slides: 79
Download presentation
CPU Scheduling

CPU Scheduling

CPU Scheduling n Basic Concepts n Scheduling Criteria n Scheduling Algorithms n Thread Scheduling

CPU Scheduling n Basic Concepts n Scheduling Criteria n Scheduling Algorithms n Thread Scheduling n Multiple-Processor Scheduling n Operating Systems Examples n Algorithm Evaluation

Objectives n To introduce CPU scheduling, which is the basis for multiprogrammed operating systems

Objectives n To introduce CPU scheduling, which is the basis for multiprogrammed operating systems n To describe various CPU-scheduling algorithms n To discuss evaluation criteria for selecting a CPU-scheduling algorithm for a particular system n To examine the scheduling algorithms of several operating systems

Basic Concepts n Maximum CPU utilization obtained with multiprogramming n CPU–I/O Burst Cycle –

Basic Concepts n Maximum CPU utilization obtained with multiprogramming n CPU–I/O Burst Cycle – Process execution consists of a cycle of CPU execution and I/O wait n CPU burst distribution

Alternating Sequence of CPU And I/O Bursts

Alternating Sequence of CPU And I/O Bursts

Histogram of CPU-burst Times

Histogram of CPU-burst Times

CPU Scheduler n Short-term scheduler selects from among the processes in ready queue, and

CPU Scheduler n Short-term scheduler selects from among the processes in ready queue, and allocates the CPU to one of them l Queue may be ordered in various ways n CPU scheduling decisions may take place when a process: 1. Switches from running to waiting state 2. Switches from running to ready state 3. Switches from waiting to ready 4. Terminates n Scheduling under 1 and 4 is nonpreemptive n All other scheduling is preemptive l Consider access to shared data l Consider preemption while in kernel mode l Consider interrupts occurring during crucial OS activities

Dispatcher n Dispatcher module gives control of the CPU to the process selected by

Dispatcher n Dispatcher module gives control of the CPU to the process selected by the short-term scheduler; this involves: l switching context l switching to user mode l jumping to the proper location in the user program to restart that program n Dispatch latency – time it takes for the dispatcher to stop one process and start another running

Scheduling Criteria n CPU utilization – keep the CPU as busy as possible n

Scheduling Criteria n CPU utilization – keep the CPU as busy as possible n Throughput – # of processes that complete their execution per time unit n Turnaround time – amount of time to execute a particular process n Waiting time – amount of time a process has been waiting in the ready queue n Response time – amount of time it takes from when a request was submitted until the first response is produced, not output (for timesharing environment)

Scheduling Algorithm Optimization Criteria n Max CPU utilization n Max throughput n Min turnaround

Scheduling Algorithm Optimization Criteria n Max CPU utilization n Max throughput n Min turnaround time n Min waiting time n Min response time

First-Come, First-Served (FCFS) Scheduling Process Burst Time P 1 24 P 2 3 P

First-Come, First-Served (FCFS) Scheduling Process Burst Time P 1 24 P 2 3 P 3 3 n Suppose that the processes arrive in the order: P 1 , P 2 , P 3 The Gantt Chart for the schedule is: P 1 0 P 2 24 n Waiting time for P 1 = 0; P 2 = 24; P 3 = 27 n Average waiting time: (0 + 24 + 27)/3 = 17 P 3 27 30

FCFS Scheduling (Cont) Suppose that the processes arrive in the order P 2 ,

FCFS Scheduling (Cont) Suppose that the processes arrive in the order P 2 , P 3 , P 1 n The Gantt chart for the schedule is: P 2 0 P 3 3 P 1 6 n Waiting time for P 1 = 6; P 2 = 0; P 3 = 3 n Average waiting time: (6 + 0 + 3)/3 = 3 n Much better than previous case n Convoy effect short process behind long process 30

Shortest-Job-First (SJF) Scheduling n Associate with each process the length of its next CPU

Shortest-Job-First (SJF) Scheduling n Associate with each process the length of its next CPU burst. Use these lengths to schedule the process with the shortest time n SJF is optimal – gives minimum average waiting time for a given set of processes l The difficulty is knowing the length of the next CPU request

Two schemes: - nonpreemptive – once CPU given to the process it cannot be

Two schemes: - nonpreemptive – once CPU given to the process it cannot be preempted until completes its CPU burst. - preemptive if a new process arrives with CPU burst length less than remaining time of current executing process, preempt. This scheme is know as the Shortest-Remaining-Time-First (SRTF). SJF is optimal – gives minimum average waiting time for a given set of processes.

Example of SJF Process Burst Arrival Time P 1 6 0 P 2 8

Example of SJF Process Burst Arrival Time P 1 6 0 P 2 8 0 P 3 7 0 P 4 3 0 n SJF scheduling chart P 4 0 P 3 P 1 3 9 P 2 16 n Average waiting time = (3 + 16 + 9 + 0) / 4 = 7 24

Example of Preemptive SJF Pid Arrival Burst Time P 1 0 7 P 2

Example of Preemptive SJF Pid Arrival Burst Time P 1 0 7 P 2 2 4 P 3 4 1 P 4 5 4 n SJF (preemptive) P 1 0 P 2 2 P 3 4 P 2 5 P 4 7 P 1 11 n Average waiting time = (9 + 1 + 0 +2)/4 = 3 16

Determining Length of Next CPU Burst n Can only estimate the length n Can

Determining Length of Next CPU Burst n Can only estimate the length n Can be done by using the length of previous CPU bursts, using exponential averaging

Prediction of the Length of the Next CPU Burst

Prediction of the Length of the Next CPU Burst

Examples of Exponential Averaging n =0 n+1 = n l Recent history does not

Examples of Exponential Averaging n =0 n+1 = n l Recent history does not count n =1 l n+1 = tn l Only the actual last CPU burst counts n If we expand the formula, we get: n+1 = tn+(1 - ) tn -1 + … +(1 - )j tn -j + … l +(1 - )n +1 0 n Since both and (1 - ) are less than or equal to 1, each successive term has less weight than its predecessor

Priority Scheduling n A priority number (integer) is associated with each process n The

Priority Scheduling n A priority number (integer) is associated with each process n The CPU is allocated to the process with the highest priority (smallest integer highest priority) l Preemptive l nonpreemptive n SJF is a priority scheduling where priority is the predicted next CPU burst time n Problem Starvation – low priority processes may never execute n Solution Aging – as time progresses increase the priority of the process

Example of Priority Scheduling Process. A arri Burst Time. T Priority Arrival P 1

Example of Priority Scheduling Process. A arri Burst Time. T Priority Arrival P 1 10 3 1 P 2 1 1 2 P 3 5 4 3 P 4 1 2 4

Round Robin (RR) n Each process gets a small unit of CPU time (time

Round Robin (RR) n Each process gets a small unit of CPU time (time quantum), usually 10 -100 milliseconds. After this time has elapsed, the process is preempted and added to the end of the ready queue. n If there are n processes in the ready queue and the time quantum is q, then each process gets 1/n of the CPU time in chunks of at most q time units at once. No process waits more than (n-1)q time units. n Performance l q large FIFO l q small q must be large with respect to context switch, otherwise overhead is too high

Example of RR with Time Quantum = 4 Process Burst Time P 1 P

Example of RR with Time Quantum = 4 Process Burst Time P 1 P 2 P 3 24 3 3 n The Gantt chart is: P 1 0 P 2 4 P 3 7 P 1 10 P 1 14 P 1 18 22 P 1 26 P 1 30 n Typically, higher average turnaround than SJF, but better response

Time Quantum and Context Switch Time

Time Quantum and Context Switch Time

Turnaround Time Varies With The Time Quantum

Turnaround Time Varies With The Time Quantum

Multilevel Queue n Ready queue is partitioned into separate queues: foreground (interactive) background (batch)

Multilevel Queue n Ready queue is partitioned into separate queues: foreground (interactive) background (batch) n Each queue has its own scheduling algorithm l foreground – RR l background – FCFS n Scheduling must be done between the queues l Fixed priority scheduling; (i. e. , serve all from foreground then from background). Possibility of starvation. l Time slice – each queue gets a certain amount of CPU time which it can schedule amongst its processes; i. e. , 80% to foreground in RR l 20% to background in FCFS

Multilevel Queue Scheduling

Multilevel Queue Scheduling

Multilevel Feedback Queue System (MFQS) n A process can move between the various queues;

Multilevel Feedback Queue System (MFQS) n A process can move between the various queues; aging can be implemented this way n Multilevel-feedback-queue scheduler defined by the following parameters: l number of queues l scheduling algorithms for each queue l method used to determine when to upgrade a process l method used to determine when to demote a process l method used to determine which queue a process will enter when that process needs service

Example of Multilevel Feedback Queue n Three queues: l Q 0 – RR with

Example of Multilevel Feedback Queue n Three queues: l Q 0 – RR with time quantum 8 milliseconds l Q 1 – RR time quantum 16 milliseconds l Q 2 – FCFS n Scheduling l A new job enters queue Q 0 which is served FCFS. When it gains CPU, job receives 8 milliseconds. If it does not finish in 8 milliseconds, job is moved to queue Q 1. l At Q 1 job is again served FCFS and receives 16 additional milliseconds. If it still does not complete, it is preempted and moved to queue Q 2.

Multilevel Feedback Queues

Multilevel Feedback Queues

Real-Time Scheduling n Can present obvious challenges n Soft real-time systems – no guarantee

Real-Time Scheduling n Can present obvious challenges n Soft real-time systems – no guarantee as to when critical real-time process will be scheduled n Hard real-time systems – task must be serviced by its deadline n Two types of latencies affect performance 1. Interrupt latency – time from arrival of interrupt to start of routine that services interrupt 2. Dispatch latency – time for schedule to take current process off CPU and switch to another

Dispatch Latency q How long it takes a real-time (RT) process to execute in

Dispatch Latency q How long it takes a real-time (RT) process to execute in the CPU once it arrives. q RT processes have the highest-priority q Dispatch latency must be small for RT process to execute fast

Dispatch Latency q Most OS force a wait for: q System call to complete

Dispatch Latency q Most OS force a wait for: q System call to complete (may be complex and long) q I/O block transfer to complete (may be too slow) before a context switch can happen

Dispatch Latency q Dispatch latency is long q Solution: place preemption points and force

Dispatch Latency q Dispatch latency is long q Solution: place preemption points and force context switch in the middle of the other process’s syscall or I/O. q. Caveat: Preemption points must be placed at “safe” locations avoid modifying kernel data structures.

Dispatch Latency What to do if RT process is waiting for resources heldby one

Dispatch Latency What to do if RT process is waiting for resources heldby one or more low-priority processes? q Priority Inversion (priority inheritance protocol) q All of them inherit the RT process’s high priority q Complete their respective tasks q Release all their resources for the RT process q All low-priority processes will revert to their original low priorities

Dispatch Latency q Dispatch latency phases: q Conflicts q Dispatch

Dispatch Latency q Dispatch latency phases: q Conflicts q Dispatch

Dispatch Latency Conflict phase: q Preempt process running in the kernel q Release low-priority

Dispatch Latency Conflict phase: q Preempt process running in the kernel q Release low-priority process’s resources needed by RT process q E. g. , in Solaris 2, q with preemption enabled: dispatch latency = 2 ms. q with it disabled: dispatch latency = 100 ms.

Dispatch Latency How long it takes a RTprocess to execute ?

Dispatch Latency How long it takes a RTprocess to execute ?

Real-Time Scheduling Process Arrival Time Burst Time P 1 0 5 P 2 2

Real-Time Scheduling Process Arrival Time Burst Time P 1 0 5 P 2 2 5 8 P 3 3 1 6 P 4 5 4 Deadline 15 12 n Earliest Deadline First/Minimum Slack (Preemptive) 0 P 3 P 2 P 1 2 3 P 2 4 P 1 P 4 8 12 n Is there another valid schedule? n What happens if P 3’s burst was 2 instead? 15

Real-Time Scheduling Process Arrival Time Burst Time P 1 0 5 P 2 2

Real-Time Scheduling Process Arrival Time Burst Time P 1 0 5 P 2 2 5 8 P 3 3 2 6 P 4 5 4 Deadline 15 12 n Earliest Deadline First/Minimum Slack (Preemptive) 0 P 3 P 2 P 1 2 3 P 2 5 P 4 P 1

Priority-based Scheduling n For real-time scheduling, scheduler must support preemptive, priority- based scheduling l

Priority-based Scheduling n For real-time scheduling, scheduler must support preemptive, priority- based scheduling l But only guarantees soft real-time n For hard real-time must also provide ability to meet deadlines n Processes have new characteristics: periodic ones require CPU at constant intervals l Has processing time t, deadline d, period p l 0≤t≤d≤p l Rate of periodic task is 1/p

Thread Scheduling n Distinction between user-level and kernel-level threads n Many-to-one and many-to-many models,

Thread Scheduling n Distinction between user-level and kernel-level threads n Many-to-one and many-to-many models, thread library schedules user -level threads to run on LWP l Known as process-contention scope (PCS) since scheduling competition is within the process n Kernel thread scheduled onto available CPU is system-contention scope (SCS) – competition among all threads in system

Pthread Scheduling n API allows specifying either PCS or SCS during thread creation l

Pthread Scheduling n API allows specifying either PCS or SCS during thread creation l PTHREAD SCOPE PROCESS schedules threads using PCS scheduling l PTHREAD SCOPE SYSTEM schedules threads using SCS scheduling.

Pthread Scheduling #include <pthread. h> #include <stdio. h> #define NUM_THREADS 5 int main(int argc,

Pthread Scheduling #include <pthread. h> #include <stdio. h> #define NUM_THREADS 5 int main(int argc, char *argv[]) { int i; pthread_t tid[NUM THREADS]; pthread_attr_t attr; /* get the default attributes */ pthread_attr_init(&attr); /* set the scheduling algorithm to PROCESS or SYSTEM */ pthread_attr_setscope(&attr, PTHREAD_SCOPE_SYSTEM); /* set the scheduling policy - FIFO, RT, or OTHER */ pthread_attr_setschedpolicy(&attr, SCHED_OTHER); /* create threads */ for (i = 0; i < NUM_THREADS; i++) pthread_create(&tid[i], &attr, runner, NULL); /* now join on each thread */ for (i = 0; i < NUM_THREADS; i++) pthread_join(tid[i], NULL); } /* Each thread will begin control in this function */ void *runner(void *param) { printf("I am a threadn"); pthread_exit(0); }

Multiple-Processor Scheduling n CPU scheduling more complex when multiple CPUs are available n Homogeneous

Multiple-Processor Scheduling n CPU scheduling more complex when multiple CPUs are available n Homogeneous processors within a multiprocessor n Asymmetric multiprocessing – only one processor accesses the system data structures, alleviating the need for data sharing n Symmetric multiprocessing (SMP) – each processor is self-scheduling, all processes in common ready queue, or each has its own private queue of ready processes n Processor affinity – process has affinity for processor on which it is currently running l soft affinity l hard affinity

NUMA and CPU Scheduling Non-Uniform Memory Access - parallel architecture where each processor has

NUMA and CPU Scheduling Non-Uniform Memory Access - parallel architecture where each processor has its own local memory but can also access memory from other processors.

Multiple-Processor Scheduling – Load Balancing n If SMP, need to keep all CPUs loaded

Multiple-Processor Scheduling – Load Balancing n If SMP, need to keep all CPUs loaded for efficiency n Load balancing attempts to keep workload evenly distributed n Push migration – periodic task checks load on each processor, and if found pushes task from overloaded CPU to other CPUs n Pull migration – idle processors pulls waiting task from busy processor

Multicore Processors n Recent trend to place multiple processor cores on same physical chip

Multicore Processors n Recent trend to place multiple processor cores on same physical chip n Each core has its own register set: seen by OS as a processor n Faster and consume less power n Multiple threads per core also growing l “Hardware” threads: hardware support includes logic for thread switching, thus decreasing the context switch time. l Takes advantage of memory stall to make progress on another thread while memory retrieve happens

Multithreaded Multicore System 0 1 2 different levels of scheduling: • Mapping software thread

Multithreaded Multicore System 0 1 2 different levels of scheduling: • Mapping software thread onto hardware thread - traditional scheduling algorithms like those discussed last time • Which hardware thread a core will run next - Round Robin (Ultra Sparc 1) or dynamic priority-based (Intel Itanium, dual-core processor with two hardware-managed threads per core)

Operating System Examples n UNIX Scheduling n Linux scheduling n Windows XP scheduling n

Operating System Examples n UNIX Scheduling n Linux scheduling n Windows XP scheduling n Solaris scheduling

UNIX schedulers

UNIX schedulers

UNIX schedulers

UNIX schedulers

Bands of Priorities

Bands of Priorities

Priority Formula

Priority Formula

Linux pre-2. 6 O(n) Scheduler O(n) scheduling a task takes O(n) where n =

Linux pre-2. 6 O(n) Scheduler O(n) scheduling a task takes O(n) where n = # of tasks One runqueue for all processors in a symmetric multiprocessor system - Task can be scheduled on any CPU - Good for load balancing - Bad for memory cache movements, e. g. , task previously on CPU 1 is run on CPU 2 move cache 1 to cache 2 Single runqueue CPUs had to contend with shared lock. No preemption allowed higher priority process may have to wait.

Linux 0(1) 2. 6 Scheduler - Each CPU has its own runqueue (priority =

Linux 0(1) 2. 6 Scheduler - Each CPU has its own runqueue (priority = 1 to 140) - Tasks are scheduled RR multilevel feedback paradigm. - Tasks whose TQ expires are moved to Expired runqueue (priorities are recalculated)

Linux Scheduling n Constant order O(1) scheduling time n Two priority ranges: time-sharing and

Linux Scheduling n Constant order O(1) scheduling time n Two priority ranges: time-sharing and real-time n Real-time range from 0 to 99 and nice value from 100 to 140

Linux 0(1) 2. 6 Scheduler Why O(1)? - Bitmap of priorities is read (each

Linux 0(1) 2. 6 Scheduler Why O(1)? - Bitmap of priorities is read (each priority level points to process) - Since size of bitmap is 140, selection of process does not depend on the number of processes in the runqueue. Active runqueue pointer - When active runqueue is empty, pointer is set to expired runqueue, i. e. , expired runqueue now becomes active runqueue and vice versa - Each CPU sets locks on its own runqueues all CPUs can schedule without contention from other CPUs.

Linux 0(1) 2. 6 Scheduler Dynamic Priority Assignment - CPU bound processes are penalized

Linux 0(1) 2. 6 Scheduler Dynamic Priority Assignment - CPU bound processes are penalized (increased by 5 levels) - I/O bound rewarded (priority# decreased by 5 levels) I/O bound use CPU to set up I/O and is suspended give other processes a chance to execute altruistic Heuristic for I/O or CPU bound category? ? - interactivity heuristic based on: time task executes compared with time it is suspended. - I/O bound sleep time is large increase in interactivity metric rewarded. - Priority adjustments only applied to user processes.

Priorities and Time-slice length

Priorities and Time-slice length

Linux Scheduling in Version 2. 6. 23 + n Completely Fair Scheduler (CFS) n

Linux Scheduling in Version 2. 6. 23 + n Completely Fair Scheduler (CFS) n Scheduling classes Each has specific priority l Scheduler picks highest priority task in highest scheduling class l Rather than quantum based on fixed time allotments, based on proportion of CPU time l 2 scheduling classes included, others can be added 1. default 2. real-time l

Linux Scheduling in Version 2. 6. 23 + n Quantum calculated based on nice

Linux Scheduling in Version 2. 6. 23 + n Quantum calculated based on nice value from -20 to +19 Lower value is higher priority l Calculates target latency – interval of time during which task should run at least once l Target latency can increase if say number of active tasks increases n CFS scheduler maintains per task virtual run time in variable vruntime l Associated with decay factor based on priority of task – lower priority is higher decay rate l Normal default priority yields virtual run time = actual run time n To decide next task to run, scheduler picks task with lowest virtual run time l

CFS Performance

CFS Performance

Linux Scheduling (Cont. ) n Real-time scheduling according to POSIX. 1 b l Real-time

Linux Scheduling (Cont. ) n Real-time scheduling according to POSIX. 1 b l Real-time tasks have static priorities n Real-time plus normal map into global priority scheme n Nice value of -20 maps to global priority 100 n Nice value of +19 maps to priority 139

Windows Scheduling n Windows uses priority-based preemptive scheduling n Highest-priority thread runs next n

Windows Scheduling n Windows uses priority-based preemptive scheduling n Highest-priority thread runs next n Dispatcher is scheduler n Thread runs until (1) blocks, (2) uses time slice, (3) preempted by higher-priority thread n Real-time threads can preempt non-real-time n 32 -level priority scheme n Variable class is 1 -15, real-time class is 16 -31 n Priority 0 is memory-management thread n Queue for each priority n If no run-able thread, runs idle thread

Windows Priority Classes n n Win 32 API identifies several priority classes to which

Windows Priority Classes n n Win 32 API identifies several priority classes to which a process can belong l REALTIME_PRIORITY_CLASS, HIGH_PRIORITY_CLASS, ABOVE_NORMAL_PRIORITY_CLASS, BELOW_NORMAL_PRIORITY_CLASS, IDLE_PRIORITY_CLASS l All are variable except REALTIME A thread within a given priority class has a relative priority l TIME_CRITICAL, HIGHEST, ABOVE_NORMAL, BELOW_NORMAL, LOWEST, IDLE n Priority class and relative priority combine to give numeric priority n Base priority is NORMAL within the class n If quantum expires, priority lowered, but never below base

Windows Priority Classes (Cont. ) n If wait occurs, priority boosted depending on what

Windows Priority Classes (Cont. ) n If wait occurs, priority boosted depending on what was waited for n Foreground window given 3 x priority boost n Windows 7 added user-mode scheduling (UMS) l Applications create and manage threads independent of kernel l For large number of threads, much more efficient l UMS schedulers come from programming language libraries like C++ Concurrent Runtime (Conc. RT) framework

Windows Priorities Priority-based preemptive scheduler: - on X axis: classes of priorities - on

Windows Priorities Priority-based preemptive scheduler: - on X axis: classes of priorities - on Y axis: relative priorities within a class Base priority for a process: threads cannot go lower Priority varies based on: - quantum used: lower priority - interrupt from keyboard: larger increase than from disk Quantum varies foreground vs. background process.

Solaris n Priority-based scheduling n Six classes available l Time sharing (default) (TS) l

Solaris n Priority-based scheduling n Six classes available l Time sharing (default) (TS) l Interactive (IA) l Real time (RT) l System (SYS) l Fair Share (FSS) l Fixed priority (FP) n Given thread can be in one class at a time n Each class has its own scheduling algorithm n Time sharing is multi-level feedback queue l Loadable table configurable by sysadmin

Solaris Dispatch Table

Solaris Dispatch Table

Solaris Scheduling

Solaris Scheduling

Solaris scheduling

Solaris scheduling

Algorithm Evaluation n How to select CPU-scheduling algorithm for an OS? n Determine criteria,

Algorithm Evaluation n How to select CPU-scheduling algorithm for an OS? n Determine criteria, then evaluate algorithms n Deterministic modeling l Type of analytic evaluation l Takes a particular predetermined workload and defines the performance of each algorithm for that workload n Consider 5 processes arriving at time 0:

Deterministic Evaluation n For each algorithm, calculate minimum average waiting time n Simple and

Deterministic Evaluation n For each algorithm, calculate minimum average waiting time n Simple and fast, but requires exact numbers for input, applies only to those inputs l FCS is 28 ms: l Non-preemptive SFJ is 13 ms: l RR is 23 ms:

Queueing Models n Describes the arrival of processes, and CPU and I/O bursts probabilistically

Queueing Models n Describes the arrival of processes, and CPU and I/O bursts probabilistically l Commonly exponential, and described by mean l Computes average throughput, utilization, waiting time, etc n Computer system described as network of servers, each with queue of waiting processes l Knowing arrival rates and service rates l Computes utilization, average queue length, average wait time, etc

Little’s Formula n n = average queue length n W = average waiting time

Little’s Formula n n = average queue length n W = average waiting time in queue n λ = average arrival rate into queue n Little’s law – in steady state, processes leaving queue must equal processes arriving, thus: n=λx. W l Valid for any scheduling algorithm and arrival distribution n For example, if on average 7 processes arrive per second, and normally 14 processes in queue, then average wait time per process = 2 seconds

Simulations n Queueing models limited n Simulations more accurate l Programmed model of computer

Simulations n Queueing models limited n Simulations more accurate l Programmed model of computer system l Clock is a variable l Gather statistics indicating algorithm performance l Data to drive simulation gathered via 4 Random number generator according to probabilities 4 Distributions 4 Trace defined mathematically or empirically tapes record sequences of real events in real systems

Evaluation of CPU schedulers by Simulation

Evaluation of CPU schedulers by Simulation

Implementation n Even simulations have limited accuracy n Just implement new scheduler and test

Implementation n Even simulations have limited accuracy n Just implement new scheduler and test in real systems n High cost, high risk n Environments vary n Most flexible schedulers can be modified per-site or per-system n Or APIs to modify priorities n But again environments vary