When Fancy Eviction Fails: Rethinking Cache Replacement For LLM Prefix Reuse

Yiyu Liu; Minlan Yu; Juncheng Yang

Home/Archive/arXiv

arXiv · 2609.28870

When Fancy Eviction Fails: Rethinking Cache Replacement For LLM Prefix Reuse

Yiyu Liu·Minlan Yu·Juncheng Yang

View original source ↗Read source document ↗Export metadata ↓

Abstract

Long-running LLM applications repeatedly send growing context, making prefix caching critical for reducing prefill cost. Yet prefix-cache behavior under agentic workloads remains poorly understood. We study production traces from two companies and evaluate 14 eviction algorithms across HBM-constrained and large memory-pool settings. Despite a large gap to Belady, sophisticated policies designed for traditional caches provide little benefit over LRU. The reason is structural: prefix reuse is dominated by the regular pacing of active sessions, making recency unusually predictive. Prefix caching nevertheless introduces new challenges, including heavy-tailed session footprints and highly variable miss costs as attention computation grows with sequence length. We introduce the compute-savings ratio and two offline oracles to quantify these effects. Our results show that effective prefix-cache management should retain recency as its foundation while selectively adding quick demotion for one-hit prefixes, compute-aware partial eviction for expensive misses, and capacity-dependent eviction granularity. We will release the traces and simulator to support future research.

Explore related subjects

cs.DC cs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Yiyu Liu, Minlan Yu, Juncheng Yang. 2026-09-24. When Fancy Eviction Fails: Rethinking Cache Replacement For LLM Prefix Reuse. https://arxiv.org/abs/2609.28870

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

The HEAL Data Platform

Objective: The objective was to develop a cloud-based, federated system to serve as a single point of search, discovery and analysis for data generated under the NIH Helping to End Addiction Long-term (HEAL) Initiative. Materials and methods: The HEAL Data Platform is built on the open source Gen3 platform, utilizing a small set of framework services and exposed APIs to interoperate with both NIH and non-NIH data repositories. Framework services include those for authentication and authorization, creating persistent identifiers for data objects, and adding and updating metadata. Results: The HEAL Data Platform serves as a single point of discovery of over one thousand studies funded under the HEAL Initiative. With hundreds of users per month, the HEAL Data Platform provides rich metadata and currently interoperates with nineteen data repositories and commons to provide access to shared datasets. Secure, cloud-based compute environments that are integrated with STRIDES facilitate secondary analysis of HEAL data. Discussion: Studies funded under the HEAL Initiative generate a wide variety of data types, which are deposited across multiple NIH and third-party data repositories. The mesh architecture of the HEAL Data Platform provides a single point of discovery of these data resources, accelerating and facilitating secondary use. Conclusion: The HEAL Data Platform enables search, discovery, and analysis of data that are deposited in connected data repositories and commons. By ensuring that these data are fully Findable, Accessible, Interoperable and Reusable (FAIR), the HEAL Data Platform maximizes the value of data generated under the HEAL Initiative.

cs.DC↗

arXiv2026

Weave: Fine-Grained Dynamic SM Scheduling in an MoE Megakernel for Compute-Communication Overlap

Mixture-of-Experts (MoE) inference under expert parallelism (EP) turns each MoE layer into a distributed computation with costly dispatch and combine communication. State-of-the-art systems reduce this cost through communication-computation overlap, splitting the GPU's SMs for communication and computation respectively. However, this approach still leaves GPU resources wasted along two dimensions. Spatially, the best SM split is determined by each layer's routing result and varies across layers and GPUs, so fixed policies mismatch the workload and waste either NVLink bandwidth or compute throughput. Temporally, complex MoE data dependencies introduce bubbles that leave SMs idle. We present Weave, to our knowledge the first MoE overlap system that performs fine-grained dynamic SM scheduling - deciding per layer and per GPU by routing results at runtime. Once routing completes, each layer's communication and computation volumes become known; Weave exploits this predictability through a lightweight cost model running inside the persistent megakernel: a spatial scheduler partitions SMs into communication workers and computation workers to match the communication/computation throughput ratio, and a temporal scheduler coordinates the two worker groups to minimize SM idleness. On 4x H100 SXM GPUs across six mainstream MoE models, Weave achieves a 2.89x geometric-mean MoE-layer speedup and a 1.33x geometric-mean end-to-end speedup over five state-of-the-art baselines.

cs.DC↗

arXiv2026

pytest-gpu-proof: Enabling Cloud-CPU Continuous Integration for GPU Code with Local GPU Attestation

GPU acceleration is now routine across robotics, but cloud-hosted GPU continuous integration (CI) runners are expensive, resulting in severe under-testing of GPU-accelerated code. We present pytest-gpu-proof, an open-source pytest plugin offering a practical middle ground. Tests can be run on a local machine, signed with a receipt of exactly what ran and what it produced, and integrated into standard CPU CI workflows (e.g., GitHub Actions). The tool is open source and on PyPI, and we are actively integrating it across our lab's software stack.

cs.DC↗