Scheduling

Overview

• problem: how to share the CPU?
• more generally: multiplex tasks onto limited resources
  • cloud service, limited machines, lots of requests
  • supermarket, limited cashiers, lots of customers
• terms/metrics
• task
• user requests, eg. mouse click, web request, shell command
• a thread can perform multiple tasks (read data, encrypt it, write to file)
• latency/response time
  • user-perceived time to finish a task (including time waiting)
• throughput
  • the rate of task completion (# of tasks done per period of time)
• scheduling overhead
  • time it takes to perform scheduling (run scheduling policy + context switch)
• fairness
  • are tasks given similar times on the CPU & wait for similar amount of time
• predictability
  • how consistent is the latency
• starvation
  • lack of progress for one task due to higher priority tasks

Scheduling Policies

• first in first out (FIFO)
  • run each task to completion, in the order they come in
  • pros/cons?
  • average latency: best case? worst case?
• shortest job first (SJF)
  • schedule the shortest job first
  • if a new shorter task arrives, preempts the current task and switches to the new task
  • pros/cons?
  • average latency?
• round robin (RR)
  • fifo but each task only gets a fixed amount of time (time slice/quantum)
  • how to pick the quantum?
  • pros/cons?
  • average latency
  • seems fair but not all tasks need CPU equally
  • some tasks use little CPU (< quantum) and blocks for events (I/O bound)
  • other tasks use more CPU and take up the full quantum (CPU bound)
• multilevel feedback queue (MLFQ)
  • multiple levels of RR queues, with increasing the quantum
  • queues with shorter quantum has higher scheduling priority
  • scheduler runs tasks within a queue in RR fashion starting from the top queue
  • when the current queue is empty, scheduler moves down to the next queue
  • when a queue with higher priority is populated with new tasks, preempt the current task and schedule tasks in the higher priority queue
  • tasks from lower queues are periodically moved up to the top priority queue to avoid starvation
  • improves latency for I/O tasks, fair to tasks that use less CPU
  • a task starts at the top queue (shortest quantum, highest scheduling priority)
  • if a task uses up the quantum, it moves down a queue (longer quantum), as it needs more CPU
  • otherwise it stays in the same queue or moves up a queue