PAPER2026-06-30 · arXiv preprint · arXiv 2607.02574

From Tensor Buffer to Distributed Memory Hierarchy: A Survey of KV Cache Management for LLM Serving

Jie Li, Tongyang Wang, Yong Chen
Compiled notes
What it moved

First systematic taxonomy of 30+ KV-cache-management systems across four design dimensions (locality, lifetime, ownership, substrate) and five architectural patterns (local-paged, disaggregated-pipeline, shared-store, memory-pool, hybrid-tier); identifies seven evaluation-methodology gaps tied to fault tolerance, tiered eviction, speculative decoding, and MoE serving. Best current map of how the serving-software layer answers the memory-wall problem this topic already tracks at the hardware layer.

KV Cache Management: From Tensor Buffer to Distributed Memory Hierarchy

The framing that matters

"The key-value (KV) cache has become a first-order memory object in LLM serving rather than a temporary per-request tensor."

That sentence is the whole argument for a serving rung. The KV cache stopped being an implementation detail inside a model and became a distributed-systems problem with its own architecture — which is why serving mechanics could not stay filed under models or hardware.

The taxonomy

Thirty-plus KV-management systems classified on four axes: locality, lifetime, ownership, substrate. Those yield five architectural archetypes — local-paged, disaggregated-pipeline, shared-store, memory-pool, hybrid-tier.

The load-bearing finding is about which axis dominates: once workload and hardware are fixed, ownership accounts for much of the remaining design variance among distributed systems. Not locality, not substrate — who owns the cache. That is a design-level claim, and it is the kind of thing a survey earns its place with.

The gap it names

The survey audits current evaluations and identifies seven missing KV-specific measurements, tying them to open problems in fault tolerance, isolation, tiered eviction, speculative decoding, MoE serving, and shared-cache semantics.

That connects straight to evals: the field is measuring KV systems without seven of the measurements it needs. A benchmark that cannot see tiered eviction cannot rank tiered-eviction designs.

Provenance

preprint, arXiv 2607.02574, submitted 2026-06-30. A survey — structure rather than a new measurement, and graded accordingly. Ingested 2026-08-06 to close a gap: this source had been listed since before the rung split with no close read, so its taxonomy was invisible to everything downstream.

Related in the base
From Tensor Buffer to Distributed Memory Hierarchy: A Survey of KV Cache Management for LLM Serving | Knowledge Base | MenFem