teleo-codex/domains/internet-finance/aimd-scaling-solves-variable-load-expensive-compute-coordination-without-prediction.md

type	domain	description	confidence	source	created	secondary_domains
claim	internet-finance	AIMD provides principled autoscaling for systems with expensive compute and variable load by reacting to queue state rather than forecasting demand	experimental	Corless et al. (SIAM 2016) applied to Teleo pipeline architecture	2026-03-11	mechanisms critical-systems

AIMD scaling solves variable-load expensive-compute coordination without prediction

For systems with expensive computational operations and highly variable load—such as AI evaluation pipelines where extraction is cheap but evaluation is costly—AIMD provides a principled scaling algorithm that doesn't require demand forecasting or optimization modeling. The algorithm operates by observing queue state: when the evaluation queue is shrinking (no congestion), increase extraction workers by 1 per cycle; when the queue is growing (congestion detected), halve extraction workers.

This approach is particularly well-suited to scenarios where:

1. Downstream operations (evaluation) are significantly more expensive than upstream operations (extraction)
2. Load is unpredictable and varies substantially over time
3. The cost of overprovisioning is high (wasted expensive compute)
4. The cost of underprovisioning is manageable (slightly longer queue wait times)

The AIMD dynamics guarantee convergence to a stable operating point where extraction rate matches evaluation capacity, without requiring any prediction of future load, modeling of arrival patterns, or solution of optimization problems. The system self-regulates through observed congestion signals (queue growth/shrinkage) and simple local rules.

The multiplicative decrease (halving workers on congestion) provides rapid response to capacity constraints, while the additive increase (adding one worker when uncongested) provides gradual scaling that avoids overshooting. This asymmetry is critical: it's better to scale down too aggressively and scale up conservatively than vice versa when downstream compute is expensive.

Evidence

• Corless et al. (2016) prove AIMD convergence properties hold for general resource allocation problems beyond network bandwidth
• The Teleo pipeline architecture exhibits the exact characteristics AIMD is designed for: cheap extraction, expensive evaluation, variable load
• AIMD's "no prediction required" property eliminates the complexity and fragility of load forecasting models
• The algorithm's proven stability guarantees mean it won't oscillate or diverge regardless of load patterns

Challenges

This is an application of proven AIMD theory to a specific system architecture, but the actual performance in the Teleo pipeline context is untested. The claim that AIMD is "perfect for" this setting is theoretical—empirical validation would strengthen confidence from experimental to likely.

---

Relevant Notes:

• aimd-converges-to-fair-resource-allocation-without-global-coordination-through-local-congestion-signals
• coordination mechanisms
• designing coordination rules is categorically different from designing coordination outcomes as nine intellectual traditions independently confirm

Topics:

• domains/internet-finance/_map
• core/mechanisms/_map
• foundations/critical-systems/_map