MADSys
The MADSys (Machine Learning, AI, Big Data Systems) group is dedicated to the design, implementation, evaluation, and application of parallel and distributed systems. Our research spans various methods aimed at accelerating data processing. Although the foundational principles like caching, batching, and overlapping are consistent, the strategic and innovative application of these techniques allows system researchers to optimally utilize diverse hardware resources across different scenarios.
Recent news
Mar 26, 2025 - Our paper “Scalio: Scaling up DPU-based JBOF Key-value Store with NVMe-oF Target Offload” is accepted by OSDI 2025！
Feb 25, 2025 - Our paper "MOONCAKE: Trading More Storage for Less Computation —— A KVCache-centric Architecture for Serving LLM Chatbot" gets 🏆FAST 2025 Best Paper!
Feb 25, 2025 - Our paper “Scaling Asynchronous Graph Query Processing via Partitioned Stateful Traversal Machines” is accepted by ICDE 2025!
Feb 25, 2025 - Our paper “OOCC: One-round Optimistic Concurrency Control for Read-Only Disaggregated Transactions” is accepted by ICDE 2025!
Feb 19, 2025 - Our opensource project “KTransformers” has earned 10k Stars⭐ at Github! Congratulations!
Sep 10, 2024 - Our paper “Transparently Share Serverless Execution Environments Across Different Functions and Nodes” is accepted by SOSP 2024!
Sep 1, 2024 - Research Assistant Professor Shan Yingdi joins MADSys Group, Welcome!
Sep 1, 2024 - Sixing Lin, Ruoyu Qin, Boxin Zhang, Jianfeng Li, Jingbo Shan, Jianwei Dong, Chen Lin, Yuanyong Chen are welcome to join the MadSys Group.
Dec 20, 2023 - Professor Wu Yongwei has been elevated to CCF Fellow. Congratulations!
Sep 15, 2023 - Wang Leping and his team participated in the “Changchengbei” cyber security competition and won the first place. Congratulations!
Sep 1, 2023 - Jinqi Hua, Xun Sun, Shaofeng Ding and Ziyu Zeng are welcome to join the MadSys Group.
Faculties
Jing Zheng
Lizhi Deng
Projects
Essentially, a decoder-only Transformer model transforms data from any modality into KVCache, positioning it as a central element in LLM serving optimizations. These optimizations include, but are not limited to, caching, scheduling, compression, and offloading. KVCache.AI is a collaborative endeavor with leading industry partners such as Approaching.AI and Moonshot AI. The project focuses on developing effective and practical techniques that enrich both academic research and open-source development.
Mooncake is the serving platform for Kimi, a leading LLM service provided by Moonshot AI.
A Flexible Framework for Experiencing Cutting-edge LLM Inference Optimizations.
Prompted by advancements in modern interconnect technologies like RDMA and CXL, this project aims to revisit the implementation and application of distributed shared memory (DSM). The objective is to facilitate the development of resilient distributed applications that can tolerate partial failures, making this process as straightforward and efficient as programming concurrent applications on a single machine. RDSM is a collaborative endeavor with leading industry partners such as Alibaba and Intel, dedicated to establishing fundamental frameworks that enhance both academic research and open-source development.
