Scheduler: Allocate CPU processors to processes over time

• Allow multiprogramming (multiple program executing on a same core)
• CPU Utilization: 1 - % of time CPU is idle
• Throughput: Work completed per unit time

Scheduler Metrics: Evaluate performance (throughput) and fairness

Short-term Scheduler / Dispatcher: Switch between ready and running threads

Process Lifecycle

Process typically alternate between CPU and I/O

• CPU-Bound: Threads with long CPU bursts and low I/O bursts
• I/O-Bound: Threads with short CPU bursts and frequent I/O bursts

Scheduling Goals

• Fairness, Avoid starvation, Policy enforcement, Balanced (all resources should be busy)

Batch Systems: Maximize throughput, minimize turnaround time, keep CPU busy

Interactive Systems: Response time (prioritize IO), proportionality (simple tasks run quickly)

Real-time Systems: Meet response deadlines, predictability

Non-preemptive Policies

Preemptive: CPU can be involuntarily taken away from a process

First-come First-serve: Always pop the thread at queue top and run to completion

• No starvation, but fair??
• Con: Convoy effect - shorter jobs waiting for one really long job

Shorter Job First: Always pop the thread with the shortest (estimated) time

• Optimal average wait time
• Con: Time estimate is inaccurate, can lead to starvation (long job waiting forever)

Preemptive Policies

Round Robin: Each process run for an equal time quantum q before switching.

• If q = ∞, same as first-come first-serve
• If q = 0, overhead of context switching wastes useful time

Priority Scheduling

Priority Scheduling: Select highest-priority job from ready queue

• Starvation: Low-P tasks may never get to run
• Priority Inversion: Low-P tasks may prevent high-P tasks from progressing
• Priority Inheritance: Tasks holding a mutex inherit the priority of task waiting on it

Single-queue: Have a ready queue sorted by priority

• Con: Sorting have too much overhead

Multi-level Queue: One ready queue per priority level

• Permanently assign priority to each process (bad)
• Each queue have its own scheduling algorithm

ML Feedback Queue: MLQ but adjust priority based on history

• Adjust priority
  • Boost aging threads waiting too long (prevent starvation)
  • Boost short threads that don’t use full quantum
  • Boost priority after user-input event
  • Lower priority for long CPU-intensive processes
• Always choose to run highest-P jobs in round-robin fashion
• High-P tasks should be shorter: IO / Interactive jobs
• Jobs start at high-P queue (increase response time)

Proportional Share Scheduling

Proportional Share: Each user gets proportional CPU time regardless of number of processes

Lottery Scheduling: Each user assigned equal “tickets,” hold a lottery to find next process

Linux Scheduling

Linux 2.6

• 140 priorities (0-99 for real-time, 100-139 for time sharing)
• O(1) scheduling
• Initial priority 120, increase priority if average sleep time is high
• Higher priority threads have longer time slice

Linux Completely Fair Scheduler

• Weighted fair queue