Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving

2024-06-24

arXiv.org

AI Summary

To address the throughput–latency SLO trade-off caused by long-context requests under high load in LLM serving, this paper proposes a KV-cache-centric decoupled architecture: it separates prefill and decoding computation across distinct resources and constructs a distributed, heterogeneous KVCache layer leveraging idle CPU, DRAM, and SSD. We design an SLO-aware, KVCache-centric scheduler that enables dynamic resource allocation and latency-prediction–guided early request rejection. Our key innovations include the first realization of decoupled KV-cache storage, coordinated scheduling across heterogeneous resources, and SLO-driven, real-time predictive scheduling. Experimental results show a 525% throughput improvement in simulation; under real-world Kimi production traffic, request capacity increases by 75% while strictly satisfying end-to-end latency SLOs.

Technology Category

Machine Learning: Large Multimodal Models (LMMs)Planning, Routing, and Scheduling: Planning with Language ModelsNatural Language Processing: (Large) Language Models

Application Category

Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved informationSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsEconomics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Mooncake is the serving platform for Kimi, a leading LLM service provided by Moonshot AI. It features a KVCache-centric disaggregated architecture that separates the prefill and decoding clusters. It also leverages the underutilized CPU, DRAM, and SSD resources of the GPU cluster to implement a disaggregated cache of KVCache. The core of Mooncake is its KVCache-centric scheduler, which balances maximizing overall effective throughput while meeting latency-related Service Level Objectives (SLOs). Unlike traditional studies that assume all requests will be processed, Mooncake faces challenges due to highly overloaded scenarios. To mitigate these, we developed a prediction-based early rejection policy. Experiments show that Mooncake excels in long-context scenarios. Compared to the baseline method, Mooncake can achieve up to a 525% increase in throughput in certain simulated scenarios while adhering to SLOs. Under real workloads, Mooncake's innovative architecture enables Kimi to handle 75% more requests.

Problem

• Optimizing LLM serving throughput under latency SLO constraints
• Managing KVCache resources in disaggregated GPU cluster architecture
• Handling request overload through prediction-based rejection policies

Innovation

• KVCache-centric disaggregated architecture separating prefill and decoding
• Leverages underutilized CPU, DRAM, SSD resources for disaggregated cache
• Prediction-based early rejection policy for overloaded scenarios