Queueing Theory Foundations

11 queueing-theory primitives for capacity planning, saturation prediction, and backpressure design. Each primitive addresses a specific failure mode that causes systems to degrade, saturate, or scale incorrectly. Primitives are domain-agnostic: the same M/M/c formula that sizes a call-center agent pool also sizes a database connection pool and a Kubernetes pod replica count.

Contents

• Quick Reference
• Primitive Index
• Formal Supporting Theory
• Misuse Boundaries
• Expert Judgment
• Decision Checklist
• Anti-Patterns
• Composition Recipes
• Workflow
• Trace-Driven Simulation
• ASCII Flow
• Related Skills
• Navigation
• Fact-Checking

Quick Reference

When to Apply

Apply queueing-theory when:

• Latency at p95/p99 grows non-linearly with load (investigate workload-specific saturation and variability)
• Queue or buffer can fill faster than it drains (request queue, message broker, thread pool)
• Capacity planning: "how many servers/replicas/workers do we need?"
• Rate-limiter or admission-control design (token bucket, leaky bucket, backpressure)
• Multi-stage pipeline where one stage's variance hurts downstream throughput

Skip and use simpler alternatives when:

• A measured latency model already meets the decision's accuracy requirement; statelessness alone does not remove queueing
• Question is about correctness under partition/failure — use foundations-distributed-systems
• Question is about reliability/availability budgets — use foundations-reliability-theory
• Question is about feedback control of a moving target — use foundations-control-theory
• Single-user dev tool with no concurrency — queueing math adds overhead with no payoff
• Measured waiting and tails are immaterial to the decision across representative load; low mean utilization alone is insufficient

Primitive Index

Formal Supporting Theory

Use references/formal-theory-map.md when the task needs stationarity, arrival-process, or distribution assumptions.

Misuse Boundaries

Check references/patterns-scenarios-traps.md before using formulas for capacity commitments.

Expert Judgment

The formulas above are correct but mechanical. What separates an expert read of a capacity problem from a formula lookup is knowing which number to distrust and why the textbook answer is usually optimistic.

Why "80% utilization" is a heuristic, not a law. No fixed utilization target follows from M/M/1. The acceptable ρ depends on the latency SLO, arrival and service variability, autocorrelation, batching, transient duration, and scaling delay. Derive a workload-specific threshold from measured traces and validate it across the operating range; published utilization bands are scenario examples, not portable targets.

VUT decomposition — variance matters as much as utilization. Kingman's formula factors cleanly into three independent levers: Variability (CV²_a+CV²_s)/2, Utilization ρ/(1−ρ), Time E[S]. When Wq blows up, an expert's first move is to ask which factor moved, not to assume it was utilization. The most common real-world regression is a variance shift with flat or even falling utilization: a new job class with a heavier tail, a noisy-neighbor GC pause, a cold-start penalty, a retry storm — all inflate CV²_s or CV²_a without moving ρ at all. Dashboards that show only "CPU 65%, looks fine" miss this entirely. If you have percentile telemetry, compare p99/p50 of service time over time — a widening ratio at flat utilization is a diagnostic clue to investigate the full second moment, workload mix and dependence; p99/p50 does not identify variance or prove which VUT factor moved, and extra capacity can reduce the U term while leaving the variance cause unresolved; compare capacity changes with variance isolation (priority lane, timeout, or separate pool).

Batch-size effects break the "μ is constant" assumption. Every formula in this skill treats service rate μ as fixed. Batching (DB writes, Kafka consumer polls, LLM continuous batching, GPU inference) makes μ a function of the current queue state — larger batches raise throughput but also raise per-item latency and effective service-time variance (a request's completion now depends on what else is in its batch, not just its own size). This is closer to a bulk-service queue (distinct from M[X]/M/1, whose X denotes batch arrivals) or a vacation-queue model than to plain M/M/1/M/G/1, and naively plugging a batch system's mean service time into P-K or Kingman can misestimate Wq in either direction because it ignores state-dependent service and correlation between co-scheduled jobs. Practical rule: if batch size is a tunable knob in the system, model it as a control variable feeding into E[S] and CV²_s, not as a constant absorbed into μ — and re-measure CV²_s at each candidate batch size rather than assuming it is batch-size-invariant.

When Little's Law is the only tool you can still trust. Every closed-form result above (M/M/1, Erlang-C, P-K, Kingman, USL) depends on distributional or stationarity assumptions — Poisson arrivals, exponential or known-moment service times, steady state, i.i.d. samples. Real production traffic routinely violates all of them at once: heavy-tailed service times where even the variance fails to converge (CV² is undefined, not just large), autocorrelated bursts from retries/cron/batch releases that a single CV²_a number cannot capture, and non-stationary regimes during incidents or autoscaling transitions. Little's Law (L = λW) is the one relationship in this skill that requires none of that — only that the system is stable and observed over a long-enough window. When you don't trust the distributional inputs a formula needs, don't force-fit Kingman or P-K anyway: fall back to measuring L, λ, and W directly and using L = λW purely as a consistency check, not as a way to derive the one unknown you can't measure. If L ≠ λW under direct measurement, the problem is measurement or population-mixing, not the formula.

Two misapplications that produce confidently wrong capacity plans:

• M/M/1 (or P-K) applied to heavy-tailed service times. When the estimated second moment is unstable across windows, P-K's Wq is not a stable capacity input. A handful of extreme requests can dominate E[S²] even when a sample CV² looks finite. Inspect tail and moment stability, then use trace-driven simulation or robust percentile analysis rather than relying on an arbitrary CV² cutoff.
• Ignoring arrival burstiness because "CV²_a looks close to 1." CV²_a measures dispersion of inter-arrival times but says nothing about correlation between them. Self-similar / long-range-dependent traffic (see Leland, Taqqu, Willinger & Wilson, "On the Self-Similar Nature of Ethernet Traffic," SIGCOMM 1993 — a foundational, widely-replicated result on bursty network traffic) can have CV²_a near 1 while still producing much longer queueing episodes than an i.i.d. renewal process with the same CV²_a, because bursts cluster in time. Kingman's formula assumes renewal (uncorrelated) arrivals and will underestimate Wq under such traffic even after "correcting" for CV²_a. If arrival autocorrelation is suspected (batch releases, coordinated retries, diurnal micro-bursts), validate against a measured autocorrelation function or a trace-driven simulation, not just a single CV²_a plugged into Kingman.

Decision Checklist

• Is the system stable? Compute ρ = λ/(c×μ). If ρ ≥ 1, no steady-state solution exists — scale capacity first.
• Single-server baseline? → M/M/1 (02). Establish the latency vs. ρ curve.
• Multiple parallel servers? → M/M/c / Erlang-C (03). Compute minimum c for wait-time SLO.
• Service time non-exponential (CV² ≠ 1)? → P-K (04) for Poisson arrivals; Kingman (07) for non-Poisson arrivals.
• Bursty arrivals (CV²_a > 1)? → Kingman (07). M/M/1 will underestimate latency.
• Multi-stage pipeline? → Jackson networks (06). Solve flow balance; find highest-ρ stage.
• Scaling horizontally? → USL (09). Fit σ and κ from load-test series; check N_max.
• Mixed SLO classes in one pool? → Priority queues (05). Separate classes; analyze each.
• High latency but good throughput? → Bufferbloat (08). Check queue depth; apply AQM or finite bounds.
• Drop-on-busy (no queue)? → Erlang-B (10). Compute blocking probability B(c, a).
• Fan-out / parallel scatter-gather? → Fork-join (11). Compute the joint maximum; use E[S] × H_K only for iid exponential service.
• Sanity-check any result? → Little's Law (01). Verify L = λ × W is consistent with measurements.
• Tail latency SLO on multi-stage pipeline with non-Poisson arrivals? → Jackson networks (06) + Ciucu-Mehri tandem sojourn bounds (SIGMETRICS 2025). Mean Jackson analysis understates tail risk when CV²_a ≠ 1.
• Does admitted work hold a growing resource until it completes (KV cache, session buffers)? → Single-resource ρ is insufficient. Check the joint compute-and-memory stability condition and the eviction/limit-cycle risk before trusting any ρ < 1 result (see Misuse Boundaries).

Anti-Patterns

Composition Recipes

Recipe 1 — Capacity Plan for a New Service

Goal: Size server pool before launch.

1. Little's Law (01): derive initial L, λ, W relationship from design requirements.
2. M/M/c (03): find minimum c so that Erlang-C wait probability meets SLO.
3. P-K / Kingman (04, 07): inflate Wq by measured CV²_s and CV²_a; re-check c.
4. USL (09): validate that the c-server pool achieves near-linear scaling (κ ≈ 0).

Standout insight: Kingman's variability factor (CV²_a + CV²_s)/2 can easily be 2–5×; a service meeting its SLO at ρ = 0.7 with M/M/c can violate SLO at the same ρ if CV² is ignored.

Recipe 2 — Saturation SLO Alert Threshold

Goal: Determine the utilization ρ* at which latency will breach SLO, and set an alert before it happens.

1. M/M/1 (02): solve W(ρ) = SLO_target; find ρ* (first-pass, exponential baseline).
2. Kingman (07): recompute ρ* with real CV²_a and CV²_s — typically ρ* is 10–20% lower.
3. Bufferbloat (08): confirm that queue depth monitoring is in place; standing queues are the first signal.
4. Little's Law (01): set alert on Lq = λ × Wq_threshold; queue depth is a leading indicator of latency breach.

Standout insight: Setting the alert on latency p99 is reactive; setting it on queue depth (via Little's Law) is proactive — the queue builds before p99 breaches.

Recipe 3 — Multi-Stage Pipeline Bottleneck Hunt

Goal: Find and fix the throughput bottleneck in a microservice chain, then verify the fix didn't shift the bottleneck.

1. Jackson networks (06): instrument each stage; collect λᵢ, μᵢ, cᵢ; solve flow-balance equations; rank by ρᵢ.
2. M/M/c (03): compute servers needed at bottleneck station i for the stated waiting/service SLO, arrival/service assumptions and uncertainty margin; choose utilization headroom from workload validation.
3. USL (09): after scaling station i, verify new ρ distribution; check for retrograde at any stage.
4. Priority queues (05): if multiple SLO classes converge at the bottleneck, separate into priority lanes.

Standout insight: The Jackson product-form result means each stage can be analyzed independently — but only after solving the traffic equations. Teams that scale one stage without re-solving flow balance routinely move the bottleneck downstream without knowing it.

Recipe 4 — LLM Inference Capacity Sizing

Goal: Size GPU/CPU capacity and select a scheduling policy for an LLM serving endpoint.

1. M/G/1 / P-K (04): model the prefill phase with Poisson arrivals and near-deterministic service (prompt-token-proportional duration); compute mean Wq only if the single-server assumptions apply; require ρ < 1 for that stationary model and validate workload-specific headroom against latency and memory targets. No universal .7 degradation threshold applies.
2. M/G/1 with SPRPT-Trail (04, 05): model the decode phase — service time = output_tokens × time_per_token, unknown at arrival. Use prediction-augmented SPRPT with Trail policy (Mitzenmacher & Shahout, Stochastic Systems 2025) to avoid KV-cache re-compute on preemption.
3. Priority queues (05): if serving multiple tiers (interactive vs. batch), apply non-preemptive priority between tiers; verify low-priority class does not starve.
4. Joint compute-and-memory stability (new, gates steps 1–3): KV cache is not a passive buffer. Each admitted request holds GPU memory that grows with every token it decodes and is released only at completion, so admitted work consumes a second resource whose demand is a function of service progress. Check the joint stability condition (Nie, Si & Zhou, ICML 2026) — a compute-side ρ well under 1 does not imply stability if memory is the binding constraint. Size the cluster from the derived stable service rate; the paper reports prediction error typically within 10% against production GPU measurements.
5. Eviction dynamics under saturation (05, 08): if the workload is homogeneous in output length, decode completions synchronize and memory peaks align, producing an evict-and-restart limit cycle with up to ~50% throughput loss (Ao, Dong, Luo & Simchi-Levi 2026). Heterogeneity is stabilizing here: coprime or dispersed decode lengths desynchronize completions. Admission-control on projected peak KV occupancy over a request's remaining decode horizon, not on instantaneous occupancy.
6. M/M/c autoscaling (03): for dynamic replica counts, apply SageServe multi-timescale control (Jaiswal et al., POMACS/SIGMETRICS 2026) — short-horizon routing + long-horizon GPU scaling via traffic forecasting.
7. Fleet simulation escalation: If tail/moment instability, batching, memory coupling or dependence makes the analytic model unreliable, combine M/G/c analytical sizing from step 1 with discrete-event simulation (inference-fleet-sim, arXiv 2603.16054) before committing to fleet purchase. Analytical M/G/c alone produces incorrect sizing for split thresholds, GPU type selection, and utilization under heavy-tailed LLM workloads.

Standout insight: Two separate throughput-optimality results now cover this setting and they answer different questions. Work-conservation is sufficient for maximum throughput on a single engine and on DAG/fork-join agent topologies (Dai, Deng, Li & Peng 2026) — which is why Orca and Sarathi-serve are throughput-optimal and vanilla vLLM is not. But work-conservation alone does not tell you how to tile prefill against decode: RAD (Bari, Hegde & de Veciana, POMACS/SIGMETRICS 2026) shows optimal tiling plus dynamic resource allocation are the binding design principles, and its SLO-aware variant SLAI cuts median TTFT 53% versus Sarathi-serve. Throughput-optimality is the floor; tiling and scheduling decide tail latency.

Workflow

1. Identify the failure mode (saturation, variance, scaling cliff, fan-out slowdown, blocking).
2. Use the Decision Checklist to map failure mode → primitive.
3. Open the primitive playbook in assets/templates/queueing-theory/ for definition, inputs, outputs, worked example.
4. For multi-failure scenarios, use the Composition Recipes or the full assets/templates/queueing-theory/README.md.
5. Validate results with Little's Law (01) — the universal consistency check.
6. Use the bundled finite-trace simulator when measured arrival/service pairs and fixed FCFS capacity are enough. Extend the event model or use a specialist simulator when finite buffers, priorities, retries, changing capacity, or shared resources matter.

Trace-Driven Simulation

Use this mode when observed burstiness or service-time variation makes a closed-form approximation too weak, while a fixed-capacity FCFS replay can answer the decision.

• Input: CSV rows with arrival_time,service_time and optional job_id; paired row values remain together.
• Command: Run from this skill directory: python3 scripts/queue_trace_simulator.py --input data/example-fcfs-trace.csv --servers 1
• Output: JSON with the explicit observation interval, busy/capacity time, utilization, queue integral and maximum, empirical wait/response quantiles, and deterministic job assignments.
• Boundary: This is one finite trace from an empty initial state. It provides no steady-state estimate or confidence interval. Use algebra when its assumptions suffice; define a validated stochastic model with warm-up, seeds, replications, and diagnostics before making population uncertainty claims.

Read references/trace-driven-simulation.md for validation rules, tie handling, the measurement denominator, output semantics, limitations, and a hand-computed example.

ASCII Flow

Waiting, capacity, or throughput problem
  -> Measure arrivals, service time, concurrency, buffer, and blocking
  -> Classify queue shape: single server, multi-server, finite buffer, priority, network, fork-join
  -> Select primitive and compute baseline
  -> Validate with Little's Law
     +-- conservation fails -> fix measurement before optimizing
     +-- conservation holds -> size capacity or simulate
  -> Return bottleneck, wait estimate, utilization risk, and scaling limit

Related Skills

Consumer skills that apply queueing-theory recipes to domain problems will reference this skill. No cross-links to non-foundation skills are made here.

Practical Decision Record

Use decision and validation worksheet for intake, model boundaries, uncertainty and checkable acceptance examples. Regression cases provide independent prompts and expected answers; these are fixtures, not executed agent results.

Navigation

• Decision and validation worksheet: references/decision-and-validation.md

• Per-primitive playbooks: assets/templates/queueing-theory/ (one file per primitive)

• Composition guide and scenario stacks: assets/templates/queueing-theory/README.md

• Formal theory map: references/formal-theory-map.md

• Patterns, scenarios, and traps: references/patterns-scenarios-traps.md

• Domain-agnostic primitives overview, anti-patterns, decision checklist: references/primitives-overview.md

• Finite-trace FCFS simulator contract and example: references/trace-driven-simulation.md

• Sources: data/sources.json

Fact-Checking

• All formulas (M/M/1, M/M/c, P-K, Erlang-B, Kingman, USL, fork-join) are derived from the primary textbook sources listed in data/sources.json. Verify against Kleinrock (1975, 1976), Harchol-Balter (2013), and Cooper (1981) before treating worked-example numbers as benchmarks.
• 2026-07-11 audit: several Harchol-Balter (2013) chapter citations in assets/templates/queueing-theory/*.md were wrong (verified against the publisher's chapter list) and have been corrected in-file with dated notes; the Kleinrock Vol.1/Vol.2 publication years were transposed in primitive 05 and are now fixed; the Erlang-C table and worked example in primitive 03, and the Erlang-B required-server table in primitive 10, contained arithmetic errors and have been recomputed and replaced with code-verified values. Treat any un-dated numeric table in this skill as a starting estimate to be recomputed, not a citation-grade constant.
• M/M/c finite-time analysis: stationary Erlang-C is not a universal lower bound on transient capacity. Initial backlog, initial server state, horizon, arrivals and target determine whether stationary sizing over- or underestimates need. An empty M/M/1 at λ=.5, μ=1 starts with zero waiting probability, versus stationary .5. Use a transient model or trace simulation; use any cited finite-time theorem only after checking its exact assumptions and bound against this initial state and target.
• USL parameters (σ, κ) are system-specific and must be fitted from load tests. Published σ/κ values for one system do not transfer to another.
• The Kingman formula is an asymptotic heavy-traffic approximation; errors increase at ρ < 0.5. Use P-K (primitive 04) for exact M/G/1 results.
• M/G/k multiserver scheduling: The prior assumption that SRPT-k achieves optimal mean response time in M/G/k queues across all loads has been superseded. Grosof & Hurtado-Lange (arXiv 2510.25963, SIGMETRICS 2026) introduce SEK-SMOD, the first policy provably achieving lower mean response time than SRPT-k across all loads and all job size distributions. Treat SRPT-k as a strong but no longer optimal baseline in M/G/k systems. Note the counterintuitive shared mechanism with the tail-latency result below: both gain by deprioritizing the shortest jobs in some regime, so "SRPT and its variants are always the right default" is no longer a safe assumption.
• M/G/k tail latency is not monotone in the policy: Yu, Harlev, Adakroy & Scully (POMACS/SIGMETRICS 2026, DOI 10.1145/3771561) prove γ-Boost is tail-constant-optimal for light-tailed M/G/k in heavy traffic, but show empirically it can be worse than plain FCFS at lighter loads. Their improved variant gives more priority to larger jobs and is both heavy-traffic optimal and stronger at light load. Practical consequence: a scheduling policy validated at peak load may be actively harmful off-peak — benchmark tail latency across the whole operating range, not only at the design point.
• 2026-08-14 audit: the citation for Dai, Deng, Li & Peng was corrected. That work retains its original title ("Throughput-Optimal Scheduling Algorithms for LLM Inference and AI Agents", arXiv:2504.07347 v3, May 2026) and is not the paper at DOI 10.1145/3771574 — that DOI belongs to a distinct POMACS/SIGMETRICS 2026 paper by Bari, Hegde & de Veciana (arXiv:2508.01002). The two results are complementary, not the same paper under a new name; both are now cited separately.
• Erlang-B and Erlang-C assume Poisson arrivals. Arrival CV²_a > 1 alone does not establish a universal bound on blocking, wait or capacity. Specify arrival dependence and burst structure; compare these stationary baselines against a fitted non-Poisson model or trace simulation.
• Fork-join E[max] = E[S] × H_K is exact for exponential service times and independent workers. Correlated sub-tasks or non-exponential service require simulation.

Learnings Loop

When prior decisions or pitfalls are relevant, consult learnings.consolidated.md if present; use learnings.md only for needed history or as the available fallback. Otherwise skip both.

After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.