Heyao Zhang

Heyao Zhang

Beijing Jiaotong University 🇨🇳

Ph.D. Student in the National Engineering Research Center of Advanced Network Technologies (ANT)

My research focuses on computing-aware networking, network for AI, AI for network, network addressing and forwarding mechanisms, and programmable data planes.

About

I am a Ph.D. student at the National Engineering Research Center of Advanced Network Technologies, Beijing Jiaotong University, advised by Prof. Hongke Zhang and Prof. Shuai Gao. My research focuses on computing-aware networking, network for AI, AI for network, network addressing and forwarding mechanisms, and programmable data planes. I am particularly interested in how future networks can efficiently coordinate network and computing resources for large-scale distributed AI applications.

I enjoy building network systems and research prototypes using programmable data-plane technologies such as P4, and exploring scalable routing and scheduling mechanisms for emerging AI-native network environments.

Research Interests

Selected Publications and Manuscripts

Srvcast: Facilitating Host-transparent and Stateful Anycast for Computing-Aware Networks
Heyao Zhang, Bo Lei, Shuai Gao, Weiting Zhang, Yuming Zhang, Cheng Chi, Hongke Zhang
IEEE International Conference on Communications 2025
CCF-C
Open-source
   [PDF]
Extended journal manuscript under review based on Srvcast
Heyao Zhang et al.
Under Review
Computing-aware Networking
Stateful Anycast
Service Routing
Manuscript under review on compute-network-aware KV cache reuse for LLM serving
Heyao Zhang et al.
Under Review
Network for AI
LLM Serving
KV Cache Reuse

Projects

Srvcast: Stateful and Scalable Anycast System for Computing-Aware Networks
Srvcast is a network-layer system designed for computation-aware networks. It enables stateful and scalable service-oriented anycast communication by integrating anycast routing, segment routing, and programmable data-plane mechanisms.

The system introduces:
  • Two-stage routing for stateful SID task steering
  • ServiceNAT for SID-to-host transparent socket compatibility
  • Selective SID advertisement for scalable routing
  • P4-based programmable data-plane implementation
[GitHub]
Srvcast system workflow
CacheRoute: Flexible KV Cache Reuse for Knowledge-Intensive LLM Serving
CacheRoute is a lightweight LLM scheduling framework built on vLLM and LMCache. It enables flexible KV cache reuse across LLM systems for knowledge-intensive services, where repeated external knowledge can be reused instead of being recomputed as long prompts.

The system introduces:
  • Knowledge-oriented task routing across LLM systems
  • KDN-based KVCache block storage and reuse
  • Dynamic selection between text-based and KVCache-based knowledge injection
  • vLLM and LMCache integration for practical LLM serving
[GitHub]
CacheRoute system design

Contact

📧 Email 23111018@bjtu.edu.cn, 1658858023@qq.com

💻 GitHub GitHub

🎓 Google Scholar Heyao Zhang Scholar Profile