HyGen: Efficient LLM Serving via Elastic Online-Offline Request Co-location

TL;DR

HyGen addresses resource underutilization in LLM inference by elastically co-locating online latency-sensitive and offline throughput-oriented workloads on a single inference engine. It introduces a latency predictor to estimate batch execution time, and an SLO-aware profiler to enforce latency budgets, paired with an interference-aware offline scheduler that employs a Prefix Sharing Maximization strategy within a two-phase, dual-queue framework. The approach achieves substantial throughput gains, up to $3.87$--$5.84\times$, while preserving strict online SLOs across multiple metrics and hardware settings. Evaluations on real production traces and diverse models demonstrate robust performance, strong predictor accuracy (MAE around $1.07$--$1.78\%$), and clear gains from prefix sharing and SLO-aware profiling, illustrating practical improvements in resource utilization for LLM serving.

Abstract

Large language models (LLMs) have facilitated a wide range of applications with distinct service-level objectives (SLOs), from latency-sensitive online tasks like interactive chatbots to throughput-oriented offline workloads like data synthesis. The existing deployment model, which dedicates machines to each workload, simplifies SLO management but often leads to poor resource utilization. This paper introduces HyGen, an interference-aware LLM serving system that enables efficient co-location of online and offline workloads while preserving SLOs. HyGen incorporates two key innovations: (1) performance control mechanisms, including a latency predictor to estimate batch execution time and an SLO-aware profiler to quantify latency interference, and (2) SLO-aware offline scheduling policies that maximize serving throughput and prevent starvation. Our evaluation on production workloads shows that HyGen achieves up to 3.9-5.8x throughput gains over online and hybrid serving baselines, while ensuring latency SLOs. The code of HyGen is publicly available at https://github.com/UIUC-MLSys/HyGen.

Figures (18)

• Figure 1: Request rate varies significantly in Microsoft Azure's LLM service over one-hour and two-minute periods.
• Figure 2: HyGen Overview. Online and offline requests are processed asynchronously, with offline requests opportunistically scheduled to respect latency budgets.
• Figure 3: HyGen respects latency requirements in co-locating requests.
• Figure 4: HyGen improves serving throughput under varying SLOs.
• Figure 5: HyGen latency predictor achieves high accuracy for batch latency prediction.