陈乐群 (Lequn Chen)

Publications

Punica: Multi-Tenant LoRA Serving

Lequn Chen, Zihao Ye, Yongji Wu, Danyang Zhuo, Luis Ceze, Arvind Krishnamurthy

arXiv, abs/2310.18547, 2023

Atom: Low-bit Quantization for Efficient and Accurate LLM Serving

Yilong Zhao, Chien-Yu Lin, Kan Zhu, Zihao Ye, Lequn Chen, Size Zheng, Luis Ceze, Arvind Krishnamurthy, Tianqi Chen, Baris Kasikci

arXiv, abs/2310.19102, 2023

Symphony: Optimized Model Serving using Centralized Orchestration

Lequn Chen, Weixin Deng, Anirudh Canumalla, Yu Xin, Matthai Philipose, Arvind Krishnamurthy

arXiv, abs/2308.07470, 2023

Nexus: A GPU Cluster Engine for Accelerating DNN-Based Video Analysis

Haichen Shen, Lequn Chen, Yuchen Jin, Liangyu Zhao, Bingyu Kong, Matthai Philipose, Arvind Krishnamurthy, Ravi Sundaram

Proceedings of the 27th Symposium on Operating Systems Principles (SOSP’19)

A Vision for Autonomous Blockchains backed by Secure Hardware

Kai Mast, Lequn Chen, and Emin Gün Sirer

Proceedings of the 4th Workshop on System Software for Trusted Execution (SysTEX’19)

AdaRes: Adaptive Resource Management for Virtual Machines

Ignacio Cano, Lequn Chen, Pedro Fonseca, Tianqi Chen, Chern Cheah, Karan Gupta, Ramesh Chandra, Arvind Krishnamurthy

arXiv, abs/1812.01837, 2018

Enabling Strong Database Integrity using Trusted Execution Environments

Kai Mast, Lequn Chen, and Emin Gün Sirer

arXiv, abs/1801.01618, 2018

Scaling Databases using Trusted Hardware Proxies

Kai Mast, Lequn Chen, and Emin Gün Sirer

Proceedings of the 2nd Workshop on System Software for Trusted Execution (SysTEX’17)

Industry Experiences

Google

Seattle, WA

Software Engineering Intern, Vertex AI Jun 2022 - Sep 2022

Disaggregated model serving.

Design and implementation a model rewriting tool for disaggregated serving.
Measurement of latency overhead and estimation of benefits in total cost of ownership.
Rule-based model optimization for disaggregated serving.

Microsoft Research

Redmond, WA

Research Intern, WatchFor Project Jun 2021 - Sep 2021

Investigated into the Pareto frontier of accuracy-latency trade-off of transfer learning models.
In-depth study of how to perform transfer learning and neural architecture search effectively and efficiently.
Explored advanced compiler optimization opportunities and challenges for the Pareto frontier models, e.g., GPU memory sharing across models, layer-based optimization caching.

Google

Kirkland, WA

Software Engineering Intern, Tango Team Jun 2019 - Sep 2019

Added a new feature to the Memcache service: replication.

Reasoned about consistency guarantees of the new feature.
Implemented with 10k lines of C++ code.
Covered by unit tests and integration tests.

In-depth discussion of the Virtual Object Set feature of the next-generation Tango.

Research Experiences

Systems Lab

Cornell University

Visiting Research Intern, advised by Emin Gün Sirer and Kai Mast Jul 2017 – Dec 2017

Worked on a database that provides blockchain-like guarantees of data integrity using Trusted Execution Environments.

Implemented large parts of the prototype on Intel SGX.
Boosted the performance inside the SGX enclave.
Increased the throughput of multi-client read workload 30x and reduced the latency by 40%.
Implemented transaction support with optimistic concurrency control.
Optimized query optimizer and executor, reducing cost of join operation to almost constant in typical workloads.
Found and solved dozens of deadlocks and data races in the initial version of the code.
Designed benchmarks and conducted experiments on a distributed testbed.

APEX Data & Knowledge Management Lab

Shanghai Jiao Tong University

Undergraduate Researcher, advised by Weinan Zhang Mar 2017 – Jun 2017

Worked on Computational Advertisement. Built a machine learning pipeline for an advertisement exchange startup.

Designed and trained a Click-Through Rate (CTR) estimation model.
Integrate the model with the startup's Real-Time Bidding (RTB) software stack.

Tianrang Network Technology Co.,Ltd

Shanghai

Research Intern, advised by Gui-Rong Xue Jun 2016 – Mar 2017

Worked on a program Yi playing board game Go similar to Google DeepMind's AlphaGo. Yi runs Monte-Carlo tree search algorithm, deep neural network, and reinforcement learning algorithms.

Designed and Implemented a distributed system running both CPU and GPU workers on multiple machines.
Reduced the network latency and increased single-machine performance.
Refactored the code base. Trained and tuned neural networks. It could beat entry-level professional human players.

Teaching Experiences

Compilers

Student Instructor Spring 2017

Led the teaching assistant team.
Redesigned assignments. Changed the target platform from SPIM MIPS Simulator to Linux x86-64, giving students more possibilities to do compilation optimization.
Built a Continuous Integration (CI) system. Once students pushed changes to their git repository, the CI would fetch the source code, compile it, test its performance on all test cases, and update the leaderboard.

Principle and Practice of Computer Algorithms

Student Instructor Summer 2016

Built an online judge system for algorithm exams.
Led a group of students to implement simplified MapReduce and Google File System. Deployed and benchmarked them on all machines of the computer room.

C++ Programming

Teaching Assistant Autumn 2015

Gave a lecture on how to begin C++ projects. Review and make comments to students' projects.
Built an online judge system for the game bot project.
Carefully designed two sets of homework to help students master the basic idea of OOP in C++.

Highlighted Projects

Compiler 2016

May 2016

Compiles C-like language to MIPS assembly, featuring code generation and optimization, including graph coloring register allocation, and transformations in Static Single Assignment form.

Isolated Deep Learning Development Environment with Shared GPUs

Jun 2017

Deep learning needs graphics cards to speed up computation. Labs sometimes cannot afford to buy graphics cards for each member. Researchers are not always good at server administration. This is a solution to isolate users while giving them the root privilege and access to GPUs without sacrificing performance.

Daily Routine Surveillance Using QQ Network Status

Jul 2014

QQ, an instant messaging software with 900 million active users, added a new feature showing users' network status (e.g. Wi-Fi/2G/3G/4G) to friends, which gave rise to privacy concerns. I invented a method so that by monitoring this status and social engineering, users' daily routine could possibly be deduced.

Simplified Git

Jun 2015

A command line source code version control system, providing the ability to archive, restore and compare source code among different versions. I implemented an efficient diff algorithm, making the speed of comparing files very fast even when the file contains tens of thousands lines of code.

Lequn Chen (陈乐群)

Research Interests

Publications

Punica: Multi-Tenant LoRA Serving

Atom: Low-bit Quantization for Efficient and Accurate LLM Serving

Symphony: Optimized Model Serving using Centralized Orchestration

Nexus: A GPU Cluster Engine for Accelerating DNN-Based Video Analysis

A Vision for Autonomous Blockchains backed by Secure Hardware

AdaRes: Adaptive Resource Management for Virtual Machines

Enabling Strong Database Integrity using Trusted Execution Environments

Scaling Databases using Trusted Hardware Proxies

Industry Experiences

Google

Microsoft Research

Google

Research Experiences

Systems Lab

APEX Data & Knowledge Management Lab

Tianrang Network Technology Co.,Ltd

Teaching Experiences

CSE550 Computer Systems

CSE552 Distributed and Parallel Systems

Compilers

Principle and Practice of Computer Algorithms

C++ Programming

Highlighted Projects

Compiler 2016

Isolated Deep Learning Development Environment with Shared GPUs

Daily Routine Surveillance Using QQ Network Status

Simplified Git