Request slowdown
Request slowdown is a job's response time divided by its positive service requirement at a specified resource.
Response time includes waiting; service requirement counts processing, including separate intervals after preemption. The ratio is dimensionless. A mean of individual ratios need not equal mean response divided by mean service. [1]
Keep the denominator explicit. FLIN's slowdown fairness uses shared versus isolated flow means; Perseus uses a workload-dependent prediction bound. Neither denominator is automatically a job's service requirement. [2]
In the synthetic FCFS example, responses are 8, 8 and 8 ms for service requirements 8, 1 and 1 ms. Mean request slowdown is 17/3, whereas the ratio of means is 12/5. [2]
Relations
- related
- Shortest remaining processing time, Service time, Slowdown fairness, Slowdown ratio, Queueing delay
Citations
- Bansal and Harchol-Balter. Analysis of SRPT Scheduling: Investigating Unfairness. SIGMETRICS 2001; sections 1 and 5.
- Hesela: a finite SRPT trace and distinct slowdown denominators.
Sources & caveats
- Definitions follow the cited queueing analysis; the numerical example uses invented arrivals and service times.
- Single-request normalization is distinct from population-level interference and detector metrics.