llm-d Open Source Inference Platform

KV-Cache-Aware Routing
Benchmark Results

Measured impact of prefix-aware scheduling on time-to-first-token latency across scaling, concurrency, and agentic workloads

3x NVIDIA H200 141GB Meta Llama 3.1 70B Red Hat OpenShift 4.21
Markell Rawls  |  AI Evangelist, Red Hat
Slide 1 of 8
Peak Performance
47.5x

Faster Time-to-First-Token at 32K Context Length

75%
TTFT Reduction
(median across scales)
73ms
P95 Latency
(cache-aware routing)
3.7s → 79ms
Cold-Start Elimination
(32K context)
2.7x -- 47.5x
Speedup Range
(500 to 32K tokens)
Slide 2 of 8
Methodology

Test Environment & Setup

Controlled single-variable comparison isolating the effect of KV-cache-aware routing on inference latency.

💻

Hardware

3x NVIDIA H200 141GB SXM GPUs, tensor-parallel across nodes

Platform

Red Hat OpenShift 4.21 with llm-d operator orchestration

Engine

vLLM v0.8.5 with prefix caching enabled, 95% GPU memory

📊

Measurement

TTFT via OpenAI-compatible streaming API, first-token timestamp delta

Test Methods

Round-Robin Baseline

Standard Kubernetes service load balancing -- requests distributed without cache awareness

Cache-Aware Routing

llm-d EPP routes requests to GPU nodes where matching KV-cache prefix blocks already reside

Controlled Variable

Only the routing strategy changes -- identical model, hardware, vLLM config, and prompt content

Slide 3 of 8
Context Length Scaling

TTFT vs. Context Length

Cold-start (round-robin) latency grows linearly with context, while cache-aware routing stays nearly flat.

Key Insight

At 32K tokens, round-robin takes 3,742ms while cache-aware completes in just 79ms -- a 47.5x improvement. The gap widens as context grows.

Why It Matters

Long-context workloads (RAG, document QA, code analysis) benefit most from cache-aware routing because their large prefill is the dominant latency cost.

Slide 4 of 8
Speedup Factor

Cache-Aware Speedup by Context Length

Ratio of round-robin TTFT to cache-aware TTFT. Higher bars indicate greater advantage from prefix caching.

2.69x
Minimum Speedup (500 tokens)
17.25x
Midrange (8K tokens)
47.45x
Maximum Speedup (32K tokens)
Slide 5 of 8
Concurrency & Distribution

Concurrent Load & Latency Distribution

How routing strategies behave under concurrent sessions and the spread of observed latencies.

Latency Distribution (8K Context)

MetricRound-RobinCache-Aware
Mean220.1 ms55.5 ms
Median234.8 ms46.3 ms
P95288.3 ms73.0 ms
Min84.9 ms41.7 ms
Max289.0 ms73.2 ms
Key Insight

Cache-aware routing maintains consistent low latency even as concurrent sessions increase. Round-robin shows high variance due to random cache misses across replicas.

Slide 6 of 8
Agentic & Fleet Scaling

Multi-Turn Agentic & Fleet Warmup Projection

How cache-aware routing performs in real agentic workflows and scales across GPU fleets.

12-Step Agentic (32K Context)

8-Step Tool Calling (8K Context)

Fleet Warmup Projection

Agentic Insight

Cache-aware routing provides stable ~70ms TTFT in 12-step agentic workflows vs. round-robin's initial 679ms spike. Multi-turn conversations benefit from prefix reuse.

Fleet Insight

Cache-aware latency stays constant at ~3.74ms (32K) regardless of fleet size, while round-robin scales linearly. At 200 replicas: 748ms vs 3.74ms.

Slide 7 of 8
Architecture

How llm-d KV-Cache-Aware Routing Works

The Prefill Problem

Every LLM request must compute KV-cache for all input tokens before generating the first output token. Longer context means longer wait.

💾

Prefix Caching

vLLM stores computed KV-cache blocks in GPU memory. If a new request shares a prefix with a cached one, it skips recomputation entirely.

🎲

Round-Robin Scatters

Standard Kubernetes load balancing distributes requests randomly, so each GPU computes the same prefix independently -- wasting GPU cycles.

🧭

llm-d Routes Smart

The Endpoint Picker (EPP) hashes prompt prefixes and routes to the GPU that already holds matching KV-cache blocks, achieving near-instant prefill.

Key Findings

Cache-aware routing delivers 2.7x to 47.5x speedup depending on context length
Benefits increase with context length -- long-context workloads see the greatest improvement
Latency remains stable under concurrent load with cache-aware routing
Agentic multi-turn workflows benefit from progressive prefix reuse
Fleet scaling shows constant latency regardless of replica count

Markell Rawls  |  AI Evangelist, Red Hat

Data collected June 2025  |  llm-d open source project

Slide 8 of 8