Resource allocation improves latency and costs in large language model services
SARA: SLO-Aware Resource Allocation for Disaggregated Agentic LLM Services
PerformanceDistributed, Parallel, and Cluster Computing
Summary
Long conversations with AI services on mobile devices can be slow due to complex computing needs. The authors of this paper created a smart system called SARA that allocates computing resources efficiently to speed up these AI services while keeping costs low. They use math models to understand different stages of AI processing and predict how to meet speed goals. Tests showed SARA improves system performance by about 27% compared to current methods without extra cost.
What this means in practice
- •For cloud infrastructure teams: Allocate computational and memory resources more efficiently to meet real-time response targets for advanced AI services while minimizing deployment costs.
- •For mobile service operators: Optimize backend AI resource use to improve latency and cost effectiveness of multi-modal agentic AI services available to mobile users.
Authors
Shicong Liu, Xianghao Yu, Zhen Gao, Jun Zhang
Abstract
Recent advances in large language models (LLMs) are driving the emergence of multi-modal and agentic services for mobile users through cloud and edge infrastructures, where long-context workloads pose daunting challenges for inference latency. Existing disaggregated LLM serving systems largely rely on hardware profiling, configuration enumeration, or heuristic scheduling, offering limited analytical guidance for cost-efficient resource allocation. In this paper, we propose SARA, a Service level objectives (SLOs)-Aware Resource Allocation framework for disaggregated agentic LLM serving systems, which maximizes goodput under a deployment cost constraint and a series of quantile-based SLO constraints. By capitalizing on queuing theory, we first model the prefill, KV cache transfer, and decode stages as an M/G/k queue, an M/G/1 queue, and a generalized birth-death process, respectively. The analysis reveals that the prefill and decode stages are dominantly limited by computational capacity and high-bandwidth memory (HBM) resources, respectively. With these mathematical models, we further derive tractable tail behaviors of different stage-wise service level metrics for both light- and heavy-tailed workloads. These characterizations explicitly map workload, model architecture, and hardware parameters to stage-wise SLO constraints and minimum resource requirements. Finally, we develop an effective resource allocation framework to maximize system goodput under limited cost budgets. Simulation and hardware results demonstrate that the proposed framework accurately predicts the stage-wise SLO with mean errors below 5%, and improves system goodput by 26.6% on average over state-of-the-art baseline methods under the same deployment cost.