---
id: wiki-2026-0508-global-network-positioning-gnp
title: Global Network Positioning (GNP)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [GNP, Network Coordinates, Network Positioning, Vivaldi]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [networking, latency-prediction, distributed-systems, p2p, coordinates]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: go
  framework: vivaldi-coords
---

# Global Network Positioning (GNP)

## 매 한 줄
> **"매 host 를 매 low-D Euclidean space 의 point 로 embed — 매 distance 가 매 network latency 의 predictor"**. Ng & Zhang (SIGCOMM 2002) 의 GNP — 매 landmark 기반 절대 좌표 — 가 매 idea 의 시초. 매 후속작 Vivaldi (Dabek 2004) 가 매 decentralized 형태로 매 P2P / overlay network 에 광범위하게 사용. 매 2026 의 application 은 매 CDN edge selection, DHT routing, server placement.

## 매 핵심

### 매 GNP (centralized) 의 idea
- **Landmark hosts** (예: 15-20개) 가 매 Internet 곳곳에 배치.
- 매 RTT 측정 후 landmark 들의 매 좌표를 매 fixed point 로 fitting (multidim scaling).
- 매 new host 는 매 landmark 들에 ping → 매 자기 좌표 solve.
- 매 두 host 간 RTT ≈ Euclidean distance.

### 매 Vivaldi (decentralized) 의 차이
- Landmark X — 매 모든 peer 가 random subset 와 매 RTT 교환.
- 매 spring relaxation: 매 over/under-estimated 매 vector 만큼 push/pull.
- 매 height augmentation (extra non-Euclidean dim) 으로 매 access link 의 last-mile 표현.
- 매 dynamic — peer churn 에 자동 적응.

### 매 application
1. **CDN routing**: 매 client 좌표 → 매 nearest edge.
2. **DHT optimization**: 매 finger table 을 매 latency-aware 로 선택.
3. **Server selection**: 매 mirror / replica 중 가장 가까운 매 fetch.
4. **Network tomography**: 매 latency map 시각화.
5. **P2P overlays**: BitTorrent, Skype 가 매 사용 (legacy).

## 💻 패턴

### GNP-style landmark fitting (NumPy)
```python
import numpy as np
from scipy.optimize import minimize

def fit_landmarks(rtt_matrix, dim=4):
    """rtt_matrix[i][j] = measured RTT between landmark i,j (ms)."""
    n = rtt_matrix.shape[0]
    x0 = np.random.rand(n * dim) * 50

    def stress(x):
        coords = x.reshape(n, dim)
        err = 0.0
        for i in range(n):
            for j in range(i + 1, n):
                d = np.linalg.norm(coords[i] - coords[j])
                err += ((d - rtt_matrix[i, j]) / rtt_matrix[i, j]) ** 2
        return err

    res = minimize(stress, x0, method='L-BFGS-B')
    return res.x.reshape(n, dim)

def position_new_host(rtts_to_landmarks, landmark_coords, dim=4):
    def stress(x):
        return sum(
            ((np.linalg.norm(x - landmark_coords[i]) - rtts_to_landmarks[i])
             / rtts_to_landmarks[i]) ** 2
            for i in range(len(rtts_to_landmarks))
        )
    res = minimize(stress, np.zeros(dim), method='L-BFGS-B')
    return res.x
```

### Vivaldi spring relaxation step (Go)
```go
type Coord struct {
    Vec    []float64 // Euclidean dims
    Height float64   // last-mile
    Err    float64   // local error estimate
}

const (
    Ce = 0.25 // error sensitivity
    Cc = 0.25 // coord change sensitivity
)

// Update local coord after measuring rtt to peer with peerCoord
func (c *Coord) Update(rtt float64, peerCoord Coord) {
    w := c.Err / (c.Err + peerCoord.Err)
    predicted := dist(c.Vec, peerCoord.Vec) + c.Height + peerCoord.Height
    es := math.Abs(predicted-rtt) / rtt
    c.Err = es*Ce*w + c.Err*(1-Ce*w)

    delta := Cc * w
    direction := unit(sub(c.Vec, peerCoord.Vec))
    force := (rtt - predicted) * delta
    for i := range c.Vec {
        c.Vec[i] += direction[i] * force
    }
    c.Height = math.Max(0, c.Height+(rtt-predicted)*delta*0.5)
}
```

### Edge selection from coordinates
```go
func selectEdge(client Coord, edges []EdgeNode) *EdgeNode {
    var best *EdgeNode
    bestRTT := math.Inf(1)
    for i, e := range edges {
        rtt := dist(client.Vec, e.Coord.Vec) + client.Height + e.Coord.Height
        if rtt < bestRTT {
            bestRTT = rtt
            best = &edges[i]
        }
    }
    return best
}
```

### CDN-style anycast hybrid (BGP + GNP fallback)
```python
def route(client_ip):
    # Try anycast result first (BGP picks PoP)
    pop = anycast_lookup(client_ip)
    if pop and recent_health(pop):
        return pop
    # Fallback: use GNP coords from RIPE Atlas / Cedexis
    coord = lookup_client_coord(client_ip)
    return min(EDGES, key=lambda e: euclid(coord, e.coord) + e.height + coord.height)
```

### Stress test (predicted vs actual)
```python
def evaluate(coords, ground_truth_rtt):
    rel_err = []
    for (i, j), actual in ground_truth_rtt.items():
        pred = np.linalg.norm(coords[i] - coords[j])
        rel_err.append(abs(pred - actual) / actual)
    return {
        'median_rel_err': np.median(rel_err),
        'p90_rel_err':    np.percentile(rel_err, 90),
    }
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Centralized infra, fixed landmarks | GNP |
| P2P / dynamic peer set | Vivaldi |
| <100 nodes | Direct measurement (skip embedding) |
| Global CDN | Anycast + GNP fallback |
| Triangle inequality violations frequent | Add height (Vivaldi) or non-Euclidean |

**기본값**: 매 modern usage — Vivaldi w/ height (4D + height).

## 🔗 Graph
- 부모: [[Distributed Systems]]
- 변형: [[Vivaldi]]
- 응용: [[CDN]]

## 🤖 LLM 활용
**언제**: 매 algorithm 설명, 매 stress function tuning 제안, embedding dimension 선택 기준.
**언제 X**: 매 real-time coord update — 매 measured RTT 만이 truth.

## ❌ 안티패턴
- **2D 만 사용**: 매 triangle inequality 자주 violated — 매 4D + height.
- **No error tracking**: 매 Vivaldi 의 local error term 빠짐 → unstable.
- **Static landmarks 의 churn 무시**: 매 landmark 매 fail 시 — health check 필수.
- **Use embedding distance for security**: 매 거리는 매 latency proxy 일 뿐 — 매 trust 의 X.

## 🧪 검증 / 중복
- Verified (Ng & Zhang SIGCOMM 2002, Dabek SIGCOMM 2004).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — GNP/Vivaldi math + edge selection patterns |