Remote caching slowdowns addressed to speed long-context AI model loading
When the KV Cache Is Slower Than the GPU: Scaling Remote KV Tiers for Context-Parallel Prefill
Distributed, Parallel, and Cluster ComputingMachine Learning
Summary
Loading memory for parts of large AI models can sometimes be slower using shared remote storage than simply recalculating the info on the GPU itself. The authors studied this problem on a powerful model with a complex memory structure and discovered that communication delays and protocol issues made remote cache lookups slower and less efficient. They improved the system by splitting the remote cache by parts, increasing parallelism, fixing bugs, and allowing asynchronous loading. This brought speed back to normal or even better, letting more tokens be served quickly from the remote cache.
What this means in practice
- •For large-scale ai deployment teams: Increase throughput and reduce latency when serving long-context language models by implementing sharded and asynchronous remote KV caches.
- •For cloud infrastructure engineers: Build remote cache tiers with parallel sharding and non-blocking loads to improve GPU utilization and reduce request queue times in disaggregated AI serving setups.
Authors
Muhammad Ibrahim
Abstract
Reusing the attention KV cache across requests is a standard way to cut prefill cost in long-context LLM serving, and production systems increasingly back the per-engine cache with a remote tier shared by every engine. We study such a tier in a disaggregated deployment of GLM-5.3, an FP8 mixture-of-experts model with multi-head latent attention (MLA), on 8xB300 nodes running prefill context parallelism (PCP8). Enabling the remote tier (LMCache with a Valkey store) left median time-to-first-token unchanged but raised the 16k-token p99 from 3.96 s to 17.3 s. Replaying a production request stream on an isolated engine, we find that a remote hit took 0.88 s against 0.31 s to recompute the prompt, and that synchronous loading turned a hit-heavy stream into an unstable queue: p99 83.32 s with the GPUs 7% busy. A two-term cost model explains both: a hit pays off only if the tier delivers more than B* = sR (2.53 GB/s for this model and GPU), and a fixed per-hit engine cost (T0 ~ 0.33 s) means that hits on prompts shorter than about 17k tokens cannot beat recompute at any bandwidth. Guided by the model, we (i) shard the MLA KV store across PCP ranks, with a proof that the existing minimum-over-ranks lookup returns the longest cached prefix and a load protocol that can never desynchronize collectives; (ii) partition the remote tier by rank, one server instance per rank with no coordination, which raises GET bandwidth 6.3x to 10.76 GB/s; (iii) add asynchronous loading to the sharded store; and (iv) fix three client bugs that caused unbounded stalls. On the hit-heavy stream, with 98.5% of prompt tokens served from the remote tier, p99 falls from 83.32 s to 0.67 s. The sharded store raises prefill throughput by 14%, and by 34% together with an upstream kernel fix. All changes are submitted upstream to LMCache.