dnsimple / erldns / 20781487450

85%

push

github

Implement smarter Congestion Control algorithms for UDP (#302)

## Overview

This PR implements the CUBIC congestion control algorithm (RFC 8312) adapted for UDP acceptors to dynamically control the admission rate based on CPU scheduler utilisation. This prevents CPU saturation during high load by intelligently throttling UDP packet acceptance. It also implements Exponential Moving Average (EMA) to the scheduler utilisation to smoothen sampling (check Nyquist-Shannon Sampling Theorem).

## Problem Statement

Under high UDP traffic, acceptors can saturate the CPU, leading to:
- Degraded performance
- Increased latency
- Potential system instability

A simple static rate limit is insufficient because:
- It doesn't adapt to actual system load
- It can't recover gracefully after congestion
- It doesn't probe for new capacity when conditions improve
- It may suffer from global synchronisation as there are multiple acceptors applying their on limiting mechanism

## Solution

### CUBIC Congestion Control Algorithm in the acceptor

The implementation adapts TCP CUBIC (RFC 8312) to control UDP acceptor admission rate:

1. **Dynamic Rate Limiting**: Rate adjusts batches/second based on system load
2. **Congestion Detection**: Monitors scheduler utilisation via `erldns_sch_mon`; triggers at 90% utilisation
3. **Smooth Recovery**: CUBIC provides graceful recovery without aggressive oscillations
4. **Capacity Probing**: Automatically probes for new capacity when system is healthy
5. **Jitter**: Adds ±7.5% jitter to sleep times to prevent synchronisation effects

- Tuned parameters:
  - `BETA = 0.8` (less aggressive than TCP CUBIC's 0.7)
  - `MAX_RATE = 1950.0` (allows 3 congestion events before engaging 1ms sleep, natural hysteresis)
  - `MIN_RATE = 1.0` (never stops completely)


### Exponential Moving Average (EMA) in the scheduler monitor

Enhanced `erldns_sch_mon` with EMA smoothing (α=0.3) for:
- More stable utilisation readings
- Reduced noise from transient... (continued)

63 of 63 new or added lines in 4 files covered. (100.0%)

1620 of 1901 relevant lines covered (85.22%)

1079.0 hits per line