Remove markdown
Lee Yunjin committed
May 6, 2026 at 10:07 UTC
c61ba0ed3c7b761a9083a2b0b06dcb4c571f9e1b
2 files changed
-69
portal/discovery/STRESS_TEST_SCENARIOS.md
deleted
-37
@@ -1,37 +0,0 @@
1
-# Extended Stress Test Specification for MOLS-EWMA Policy
2
-
3
-This document outlines long-term stress test scenarios designed to validate the stability, scalability, and transposition accuracy of the EWMA-based relay selection policy.
4
-
5
-## 1. Scenario A: "The Messy Grid" (64x64 Non-Ideal Distribution)
6
-**Objective:** Validate transposition logic when relay density is non-optimal (not a perfect divisor of 64), which naturally occurs in fragmented network topologies.
7
-
8
-- **Setup:** 53 Relay URLs (not a power of 2, creating uneven GF(64) hash mapping).
9
-- **Network Stress:**
10
- - Inject "Micro-burst" RTT spikes (100ms duration, 800ms magnitude) every 30 seconds to 30% of nodes.
11
- - Validate that EWMA ($\alpha=0.3$) filters these spikes, preventing frequent priority flapping.
12
-- **Success Criteria:**
13
- - The system must prioritize the remaining 70% of stable nodes.
14
- - No "Priority Oscillation" where a node bounces between rank 1 and 5 every minute.
15
-
16
-## 2. Scenario B: "Massive Scale" (256 Node Pool)
17
-**Objective:** Validate performance and memory stability when the relay set significantly exceeds the standard 64-node MOLS grid.
18
-
19
-- **Setup:** 256 active Relay URLs.
20
-- **Network Stress:**
21
- - Perform a "Rolling Congestion" simulation: shift a 500ms+ latency penalty across groups of 32 nodes sequentially over a 24-hour period.
22
-- **Success Criteria:**
23
- - **Latency:** `rankRelayPool` execution time must remain < 5ms under the increased load.
24
- - **Stability:** The transposition logic should maintain a consistent set of the "top 3" healthiest nodes even as congestion rolls across the 256-node pool.
25
- - **Memory:** Telemetry counters (Prometheus gauge tracking) must not grow beyond the defined `boundedRelay` limit of 1024.
26
-
27
-## 3. Implementation Guidelines for `portal-loadtest`
28
-- **Simulation Duration:** All scenarios should run for a minimum of 24 hours to observe EWMA convergence.
29
-- **Telemetry Hook:** Integrate with the `portal/telemetry` package to log the `chi-square` uniformity metric every hour alongside the `EWMA RTT` distribution.
30
-- **Command:**
31
- ```bash
32
- # Scenario A
33
- make load-test -- -clients 500 -relays 53 -mode messy
34
-
35
- # Scenario B
36
- make load-test -- -clients 2000 -relays 256 -mode scale
37
- ```
portal/discovery/TEST_MOLS.md
deleted
-32
@@ -1,32 +0,0 @@
1
-# MOLS EWMA & Percentile Transposition Performance Validation
2
-
3
-## 1. Overview
4
-We have upgraded the relay selection policy to incorporate not only EWMA-smoothed RTT but also percentile-based jitter analysis (p99 - p1). This multi-layered approach ensures that nodes are prioritized based on both central tendency (stability) and consistency (predictability).
5
-
6
-## 2. Methodology
7
-- **Percentile Tracking:** Each relay tracks the last 100 RTT samples.
8
-- **Jitter Scoring:** We calculate `Jitter = p99 - p1`.
9
-- **Transposition Criteria:** A relay is demoted if:
10
- - `EWMA RTT > 500ms` (persistent congestion) OR
11
- - `Jitter > 200ms` (high inconsistency/predictability risk)
12
-
13
-## 3. Performance Metrics (Simulated vs. Expected)
14
-
15
-| Metric | Threshold | Logic | Impact |
16
-| :--- | :--- | :--- | :--- |
17
-| **p50 (Median)** | < 100ms | Primary selection | Baseline low-latency path. |
18
-| **p99 (Tail)** | < 500ms | Transposition trigger | Prunes transient congestion spikes. |
19
-| **Jitter (p99-p1)**| < 200ms | Consistency filter | Eliminates "unpredictable" nodes. |
20
-
21
-## 4. Test Results (from `TestMOLSSelectPriorityEWMAStabilityTransposition`)
22
-The transposition logic was verified under a simulated scenario comparing a stable node against an inconsistent/high-jitter node.
23
-
24
-- **Stable Relay:** `EWMA=100ms`, `Jitter=20ms` -> **Ranked #1**
25
-- **Unstable Relay:** `EWMA=600ms`, `Jitter=300ms` -> **Ranked #2 (Demoted)**
26
-
27
-The engine successfully correctly identified and demoted the unstable node, even when their base MOLS scores were mathematically equivalent.
28
-
29
-## 5. Expected Performance Gains
30
-1. **Selection Predictability:** By penalizing nodes with high jitter, we steer traffic toward nodes that offer a tighter latency distribution, reducing re-transmission rates and improving throughput consistency.
31
-2. **Jitter Resilience:** The use of `p99 - p1` spread proactively identifies nodes subject to path oscillation or bufferbloat before they fully degrade the active session.
32
-3. **Tail Latency:** The combined EWMA and percentile filtering is expected to reduce p99 latency by **20-30%** compared to the original purely-MOLS-based policy.