Inside a Trillion-Token/Day Factory: Mooncake Turns KVCache into a Cluster-Shared Pool
量子位 · wechat · 2026-09-11
QuJing/KVCache.AI details the inference architecture behind a trillion-token-per-day production system, where per-machine token efficiency rose 3x and total capacity 30x since Spring 2026 — powered by pooling KVCache across the cluster with MooncakeStore.
- Agentic workloads hammer long-context reuse; a single cache miss can force re-prefill of hundreds of thousands of tokens
- Architecture: Prefill-Decode separation, SGLang HiCache on Prefill only, MooncakeStore as an independent service pooling host DRAM, RDMA over 800Gbps NICs, engine/cache decoupled
- SMG-based cluster scheduling balances cache locality vs. hotspots; RBG orchestrates roles on Kubernetes
- Cache hit rate holds above 90%; average batched KVCache reads under 50ms via async prefetch overlapping compute
- War stories: LRU eviction's shard write-lock contention fixed by object-granular locks and moving slow ops out of critical sections; node scale-up/down accelerated several-fold via memory/RDMA registration optimizations and huge pages
More from Infra
- Routing NVIDIA PAIR to llama.cpp on an AMD ROCm node (2×R9700): full notes — Don_Reuter · 2026-09-11
- Running Qwen3.8 locally on a 128GB laptop for agentic coding: thinking tokens, not tok/s, set the wall clock — deepu105 · 2026-09-11
- SpaceX plans to make scarce turbine parts as AI data centers outpace gas turbine supply — rohanpaul_ai · 2026-09-11
- AGI as task time horizon vs meetings — and why fabs should train their own models — jwt0625 · 2026-09-11
- Free client-side calculator compares LLM token economics across DeepSeek, Claude, o3-mini — nikola_mr64990 · 2026-09-11
- 10 resources on what happens after training: KV-cache, quantization, serving — techNmak · 2026-09-11