Slide 1 of 8
Live Cluster Simulation

Transparent Cluster
Observability

Inside the llm-d production inference cluster. Real-time metrics, GPU topology, intelligent routing, and full-stack tracing for LLM serving at scale.

96
Health Score
4 / 4
Pods Healthy
8
H100 80GB GPUs
99.97%
Uptime (30 days)
Slide 2 of 8

The Cluster at a Glance

A DGX H100 node running 4 inference pods serving Llama 3.1 70B with live LoRA adapter hot-swapping, prefix-aware caching, and sub-second time-to-first-token.

Llama 3.1 70B
Base Model
NVLink 4.0
450 GB/s per link
RDMA RoCE v2
Network Fabric
3 Active
LoRA Adapters
lora-finance-v2
pod-0
lora-code-v1
pod-1
lora-medical-v3
pod-2
Slide 3 of 8

GPU Topology: DGX H100

Two NUMA domains, each with 4 H100 80GB GPUs connected via NVLink 4.0, bridged by NVSwitch at 450 GB/s bidirectional throughput.

Slide 4 of 8

Real-Time Performance

Continuously streamed metrics across all inference pods. From time-to-first-token latency to prefix cache hit ratios.

182ms
TTFT p50
3,240
Tokens/sec
68%
Prefix Cache Hit
0.12%
Error Rate
Slide 5 of 8

Intelligent Routing with EPP

Every request entering the Inference Gateway (IGW) is scored by the Endpoint Picker (EPP) using KV-cache-aware, load-aware scheduling. Prefix-cache affinity, current load, queue depth, and pod health all factor in. Weights below are illustrative.

0.40
Cache Affinity
0.25
Load Balance
0.20
Queue Depth
0.15
Health Factor
Slide 6 of 8

Request Lifecycle Tracing

Every inference request is traced through six stages. From gateway ingress through EPP routing, prefill, KV transfer (when prefill and decode are disaggregated), token decode, and response delivery.

Slide 7 of 8

When Things Go Wrong

The cluster detects anomalies in real time. Latency spikes, GPU OOM events, pod crashes, and cache storms, with structured remediation guidance for every alert.

CRITICAL: TTFT p99 Spike
TTFT p99 spike detected: >8000ms across cluster. Check KV cache eviction rates. Consider reducing max batch size.
WARNING: Cache Storm Subsiding
Prefix cache is rebuilding. Full recovery expected in 2-5 minutes depending on request distribution.
RESOLVED: Pod Recovered
vllm-pod-2 recovered from OOM event. Monitor cache hit rate recovery. Pod is accepting requests again.
kubectl logs -f deployment/vllm --all-containers
Slide 8 of 8

Observability You Can Trust

llm-d brings transparent, production-grade observability to LLM inference. From GPU silicon to application response. Every layer visible. Every decision traceable. Every anomaly actionable.

GPU Topology
Full NUMA-aware view of 8x H100s, NVLink mesh, thermal state
Intelligent EPP
Load-aware prefix-cache routing with real-time scoring
Full Tracing
End-to-end request lifecycle from gateway to token decode