RT Policies: SCHED_FIFO vs. SCHED_RR vs. SCHED_DEADLINE

The Three Flavors of Real-Time

1. SCHED_FIFO (First-In, First-Out)
The simplest and most aggressive RT policy.

• A SCHED_FIFO process runs until it either blocks (e.g., waits for I/O), voluntarily yields (sched_yield()), or is preempted by a higher-priority RT task.
• There is no time slice. If a SCHED_FIFO priority-50 process is runnable, it runs forever (subject to RT throttling) until one of the above conditions occurs.
• A bug causing an infinite loop in a SCHED_FIFO process will consume the CPU entirely (minus the RT throttle reserve).
• Use case: Interrupt handlers, audio servers (like PulseAudio/PipeWire), industrial control loops.

2. SCHED_RR (Round-Robin)
A gentler version of FIFO.

• Identical to SCHED_FIFO except it has a time quantum (default 100ms, visible in /proc/sys/kernel/sched_rr_timeslice_ms).
• When a SCHED_RR process exhausts its quantum, it moves to the back of the queue of RT tasks at the same priority level. Higher-priority RT tasks still preempt it immediately.
• Use case: Multiple RT tasks at the same priority that should share the CPU fairly among themselves.

3. SCHED_DEADLINE
The most sophisticated RT policy, implementing Earliest Deadline First (EDF) scheduling.

• Each task declares three parameters: runtime (how much CPU it needs per period), deadline (by when it must finish), and period (how often it repeats).
• Example: "I need 2ms of CPU every 10ms, and I must finish within 8ms of my period start."
• The kernel performs admission control: it will reject a SCHED_DEADLINE task if accepting it would make existing deadline tasks miss their deadlines.
• SCHED_DEADLINE tasks have higher priority than SCHED_FIFO and SCHED_RR.
• Use case: Video encoding pipelines, network packet processing, any workload with known, periodic CPU requirements. In practice, these three parameters are set on a running process via chrt --deadline, which takes the runtime, deadline, and period as arguments (in nanoseconds).