---
type: claim
domain: internet-finance
description: "Self-correcting resource allocation without forecasting by responding to current system state"
confidence: likely
source: "Vlahakis, Athanasopoulos et al., AIMD Scheduling and Resource Allocation in Distributed Computing Systems (2021)"
created: 2026-03-11
---

# AIMD scaling policy eliminates load prediction requirements by reacting to observed queue state through additive increase when stable and multiplicative decrease when congested

AIMD provides an elegant solution to the resource allocation problem: when queue is shrinking (system healthy), add workers linearly (e.g., +1 per cycle). When queue is growing (system overloaded), cut workers multiplicatively (e.g., halve them). This reactive approach eliminates the need to predict future load because the system continuously adjusts based on current observed state.

The policy is self-correcting: over-allocation naturally corrects through queue shrinkage triggering slower growth, while under-allocation corrects through queue growth triggering aggressive scaling. The asymmetry (additive increase, multiplicative decrease) prevents oscillation while maintaining responsiveness.

Vlahakis et al. prove this approach achieves stable resource allocation regardless of system scale or parameter choices, making it robust to implementation variations and deployment contexts.

## Evidence

- AIMD responds to queue state (shrinking vs growing) rather than attempting to forecast load patterns
- Linear increase during stability prevents over-reaction to noise
- Multiplicative decrease during congestion provides rapid response to capacity constraints
- Proven stable irrespective of total node count and AIMD parameters
- The mechanism requires only local queue observation, not global system state or historical load data

## Application to Pipeline Architecture

For systems like Teleo's extract-eval pipeline, AIMD means: when the eval queue is shrinking, add eval workers at constant rate. When the eval queue is growing, cut eval workers by a fixed percentage. This handles variable extract rates without requiring extract throughput forecasting or complex capacity planning.

The connection to Little's Law (mentioned in the source) enables QoS guarantees: local queuing time is calculable from simple formulas given AIMD parameters, allowing latency bounds without centralized scheduling.

---

Relevant Notes:
- [[aimd-congestion-control-generalizes-to-distributed-resource-allocation-because-matching-producer-rate-to-consumer-capacity-is-structurally-identical-across-networks-and-compute-pipelines]] <!-- claim pending -->
- [[coordination mechanisms]]

Topics:
- core/mechanisms/_map
- domains/internet-finance/_map