Mooncake / Dynamo mapping

QuillCache replicates the production reference designs piece by piece, then adds its differentiation on top. The concepts line up one-to-one — built small enough to read end-to-end and measure.

Mooncake / NVIDIA Dynamo	QuillCache	Status
Mooncake Store (pooled DRAM/SSD KV)	LocalKvStore + PooledStore	✅ real bytes
Mooncake Transfer Engine	quillcache-transfer	✅ TCP / ⊙ RDMA reserved
Conductor / scheduler	quillcache-control + router	✅
Dynamo KV-router cost function	DynamoCostRouter	✅ reproduces the worked example
Dynamo KVBM tiers (G1/G2/G3)	StoreDataPlane (HBM/DRAM/SSD)	✅ moves real bytes
Dynamo KV-Cache Indexer	residency index (Holt ART)	✅ persistent
Dynamo etcd / service discovery	NodeRegistry (StaticRegistry)	✅ etcd pluggable
— (neither does this)	identity guard + crash-consistency	🎯 differentiation

The Dynamo cost function

DynamoCostRouter reproduces the cost function NVIDIA Dynamo’s KV router runs. For each worker:

overlap_credit   = 1.0·device + 0.75·host + 0.25·disk   (HBM / DRAM / SSD hits)
adjusted_prefill = max(0, raw_prefill_blocks − overlap_credit)
cost             = prefill_load_scale · adjusted_prefill + decode_blocks

It routes to the lowest-cost worker. A GPU-resident prefix hit is worth 4× an SSD one — cache locality vs load, as a single number — and a unit test reproduces Dynamo’s own published worked example (costs 18 / 10 / 11 → pick worker 2).

The distributed read path

The pooled read mirrors Mooncake’s Conductor → metadata → Transfer Engine flow:

1. the residency index locates which nodes hold the block (index.locate);
2. the NodeRegistry resolves a node id to its transfer address;
3. the transfer engine fetches it; the block is admitted locally and served, identity-guarded.