LLM Serving Optimization with Variable Prefill and Decode Lengths
This addresses a critical bottleneck in LLM serving for real-world applications by optimizing memory usage and scheduling efficiency, though it is an incremental improvement over existing scheduling methods.
The paper tackles the problem of scheduling LLM requests with varying prefill and decode lengths to minimize total completion time, proving that common strategies like FCFS and SF have suboptimal competitive ratios and proposing a novel algorithm with a constant competitive ratio that outperforms baselines in simulations.
We study the problem of serving LLM (Large Language Model) requests where each request has heterogeneous prefill and decode lengths. In LLM serving, the prefill length corresponds to the input prompt length, which determines the initial memory usage in the KV cache. The decode length refers to the number of output tokens generated sequentially, with each additional token increasing the KV cache memory usage by one unit. Given a set of n requests, our goal is to schedule and process them to minimize the total completion time. We show that this problem is NP-hard due to the interplay of batching, placement constraints, precedence relationships, and linearly increasing memory usage. We then analyze commonly used scheduling strategies in practice, such as First-Come-First-Serve (FCFS) and Shortest-First (SF), and prove that their competitive ratios scale up sublinearly with the memory limit-a significant drawback in real-world settings where memory demand is large. To address this, we propose a novel algorithm based on a new selection metric that efficiently forms batches over time. We prove that this algorithm achieves a constant competitive ratio. Finally, we develop and evaluate a few algorithm variants inspired by this approach, including dynamic programming variants, local search methods, and an LP-based scheduler, demonstrating through comprehensive simulations that they outperform standard baselines while maintaining computational efficiency.