Three questions, one tension, and conserved work
Which three questions does scheduling answer, and why can ordering only move delay, never create capacity?
The three questions of scheduling
This mechanism has three subsections. They answer three different questions.
| Subsection | Question | One-line description |
|---|---|---|
| Job and task scheduling | When? | Ordering in time over a fixed pool of resources |
| Placement | Where? | Matching work to machines that are not uniform and not uniformly priced |
| Workload management | Whether, and how much at once? | The control loop that keeps the first two problems solvable |
The dependency runs in one direction. Ordering redistributes delay but conserves it. So no scheduling policy can rescue a system that is admitted past its capacity. Admission control (workload management) is therefore not a sibling of scheduling. It is a precondition for it. That is why the third subsection exists, and why most production "scheduler problems" are actually admission problems.
The central tension: isolation versus utilization
One tension appears on almost every page of this mechanism:
A tenant is any user, team or application sharing the system. If you promise Omar's team 30% of the cluster at any moment, then 30% must be free (or quickly freeable) even when Omar's team is asleep. That idle 30% is the price of the promise. Other tensions appear in smaller form: batching decisions (gangs), static slots versus adaptive admission, rejecting versus queueing, and the cost of placement choices.
Work is conserved: only delay moves
Here is the arithmetic fact on which this whole mechanism rests.
A work-conserving policy never leaves a resource idle while work is waiting. For any work-conserving policy, the total busy time is the same. So throughput and utilization under saturation do not depend on the policy. What the order does change is the mean flow time.
- when job i finishes
- mean time from arrival to completion
The total work is the same (101 minutes). The machine is busy the whole time in both cases. But the mean wait is about double when the long job goes first. All of scheduling lives in that gap.
Now see why Lena's priority flag did nothing. If everyone is "high," the order is the same as before. And even a perfect ordering cannot change the fact that Hub is asked to do more work than 200 machines can finish. Ordering only chooses which users wait.