Queue fairness and aging

FIFO orders arrival times, static priority favors configured classes, and fair-share policies compare consumption or entitlement across identities. None of these descriptions completely specifies a scheduler. A strict head-of-line job can block feasible work behind it, while backfilling can improve utilization if it does not invalidate a protected future start under the supported assumptions. A stream of small jobs can prevent a large gang from ever accumulating enough free resources unless the policy reserves or reclaims a suitable group.

Define the identity receiving fairness: job, user, team, project, or service class. Fairness by job can reward a tenant for splitting one workload into many submissions. Report wait distributions by identity and shape, including rejected and cancelled work rather than only admitted jobs. Separate time waiting for quota from time waiting for physical placement or application readiness. A job whose requirements exceed the entire eligible cluster needs a clear infeasibility outcome, not indefinite promotion through a queue that can never satisfy it.

An increasing score can help old work outrank new work within its allowed class. A cap may deliberately preserve class separation but can also leave a lower-priority job permanently behind an endless higher-priority stream. An uncapped score changes those tradeoffs and may conflict with urgent-service expectations. State whether aging survives retries, preemption, or user edits; otherwise a job can lose accumulated waiting credit or manipulate its position. Keep score components visible enough for users to understand a decision without revealing other tenants’ sensitive details.

Fairness and job-completion time are different objectives. A policy based on attained service can favor jobs that have consumed little so far without requiring exact remaining runtime, but that choice has assumptions and starvation countermeasures. Tiresias studies this setting for distributed deep learning. Do not interpret a paper’s improvement on its evaluated traces as proof that an arbitrary aging coefficient benefits your fleet. Large-job shape, placement locality, runtime uncertainty, and preemption overhead affect outcomes beyond queue rank.

When sustained admitted demand exceeds the resources that can complete work, no ordering policy can give every job a fixed finite wait without changing admission or capacity assumptions. Combine fairness with quotas, overload handling, feasible reservation, or scaling. Watch the oldest eligible wait and large-gang progress alongside average completion metrics. A scheduler can improve an average by indefinitely postponing a small number of expensive jobs, which is a policy choice that should be visible rather than hidden inside a favorable headline.

The queue lab demonstrates only the capacity condition. Reset, set arrival rate to 4 and service rate to 3, predict backlog after one second, then Step twice. For a separate run, Reset and explicitly set arrival rate to 2 and service rate to 3, predict, then Step twice. Both begin with four equal abstract jobs. The fluid slope can show overload, but there is no priority, aging, tenant identity, or variable job size. Explain why changing queue order alone cannot reverse its rate imbalance.