Scalable and Fast Inference Serving via Hybrid Communication Scheduling on Heterogeneous Networks

Advances in large language models (LLMs) have opened up new possibilities across various fields, fueling a new wave of interactive AI applications such as DeepSeek and ChatGPT. Inference serving systems play a crucial role in supporting these applications. Recent research indicates that when crossserver parallelization is enabled in inference serving systems, data synchronization overhead can exceed 65% of the total inference delay, making the reduction of communication overhead essential for speeding up inference. While existing systems accelerate cross-server communications by offloading synchronization operations to programmable switches, they often suffer from limited aggregation throughput under bursty traffic conditions, posing challenges for homogeneous network environments. To address these challenges, we propose HeroServe, an innovative inference serving system that leverages heterogeneous networks to accelerate data synchronization in distributed clusters. Our approach enables a fast and scalable inference serving system by employing an offline planner for joint computation allocation and communication scheduling, along with an online scheduler for dynamic traffic management and load balancing. We implement a prototype on a testbed comprising six servers and two programmable switches. Experimental results demonstrate that HeroServe improves scalability by $1.53 \times$ while achieving lower latency compared to state-of-the-art solutions.