🛰️ Daily AI Frontier
‹ back to 2026-08-22

CacheRoute: Planned Prefix-Affinity Routing for Large-Scale LLM Serving

Research Efficiency & Systems

Ranking

Overall 88
Content 95
Popularity 71

Observed public metrics from 1 member.

Merged summary

TL;DR - CacheRoute periodically plans prefix-affinity routing so repeated LLM requests reach servers holding reusable KV caches without severely skewing load. On Llama-3.3-70B across 60 H100 GPUs, it delivered 2.3× the throughput of the strongest baseline under a 3.5-second p99 latency SLO.

  • CacheRoute assigns high-rate keys to stable warm server sets and places them according to expected load; exceptionally hot keys may use multiple destinations.
  • It sustained 176±11 QPS and increased served KV-cache hit rate from 64.1±1.3% with cache-blind balancing to 93.2±0.5%.
  • Experiments isolate gains from both request affinity and load-aware placement across multiple model sizes and traffic patterns.
  • Two 32B workloads showed that insufficient KV reuse can let residual load skew reduce or eliminate gains, motivating shadow-replay validation before deployment.

Sources (1)

CacheRoute: Planned Prefix-Affinity Routing for Large-Scale LLM Serving

arXiv cs.DC Huang Cheng 2026-08-20 arXiv:2608.19677
Public signals Semantic Scholar citations 1 · Semantic Scholar influential citations 0
Providers: Hugging Face · N/A OpenAlex · N/A Publisher · N/A Semantic Scholar · Citations 1 · Influential citations 0 X · N/A Fetched 2026-09-14 14:19:07.131725 UTC

TL;DR - CacheRoute periodically plans prefix-affinity routing so repeated LLM requests reach servers holding reusable KV caches without severely skewing load. On Llama-3.3-70B across 60 H100 GPUs, it delivered 2.3× the throughput of the strongest baseline under a 3.5-second p99 latency SLO.

  • CacheRoute assigns high-rate keys to stable warm server sets and places them according to expected load; exceptionally hot keys may use multiple destinations.
  • It sustained 176±11 QPS and increased served KV-cache hit rate from 64.1±1.3% with cache-blind balancing to 93.2±0.5%.
  • Experiments isolate gains from both request affinity and load-aware placement across multiple model sizes and traffic patterns.
  • Two 32B workloads showed that insufficient KV reuse can let residual load skew reduce or eliminate gains, motivating shadow-replay validation before deployment.
item →