Non-preemptive priority
Non-preemptive priority selects higher-priority pending work at scheduling boundaries without interrupting the operation already in progress.
A two-class node with strict aggregate service beta and maximum lower-priority packet work L gives the high class beta_H(t)=max(0,beta(t)-L). For beta(t)=R*max(0,t-T), the added latency is L/R. [1]
In a hypothetical serial server delivering 1 MiB/ms, an 8 MiB low-priority command starts at time zero. A 1 MiB high-priority request arrives at 0.5 ms: it waits 7.5 ms and completes 8.5 ms after arrival. Four 2 MiB commands with priority reconsidered between them reduce those times to 1.5 and 2.5 ms. Splitting bytes without exposing a scheduling boundary does not produce that result. [4]
Seek, rotation and internal work can dominate payload transfer. Semi-preemptible I/O exposes command boundaries, and actual interruption can add positioning cost. [2]
Linux I/O-priority support depends on the scheduler; its realtime class can starve other work. Selecting that class alone supplies neither a service curve nor an end-to-end deadline. [3]
Relations
Citations
- Le Boudec and Thiran. Network Calculus, 2022 online edition, Corollary 6.2.1, p. 176.
- Dimitrijevic, Rangaswami and Chang. Design and Implementation of Semi-preemptible IO. FAST 2003, sections 1, 3 and 4.
- Linux 6.16 documentation: Block I/O priorities, introduction and scheduling classes.
- Hesela: exact hypothetical non-preemptive priority traces and bounds.
Sources & caveats
- The service bound is a published theorem for a specified node, not an empirical SSD guarantee. The trace is exact invented arithmetic, not a reproduction of FAST 2003.
- FAST 2003 studied a historical SCSI disk prototype. Device parallelism, firmware queues and retries require their own work and scheduling contracts.