It Takes Two to Entangle

Zhanghan Wang*, Ding Ding*, Hang Zhu, Habin Lin, Aurojit Panda.

Understanding Stragglers in Large Model Training Using What-if Analysis

Jinkun Lin, Ziheng Jiang, Zuquan Song, Sida Zhao, Menghan Yu, Zhanghan Wang, Chenyuan Wang, Zuocheng Shi, Xiang Shi, Wei Jia, Zherui Liu, Shuguang Wang, Haibin Lin, Xin Liu, Aurojit Panda, and Jinyang Li.

Runtime Protocol Refinement Checking for Distributed Protocol Implementations

Ding Ding, Zhanghan Wang, Jinyang Li, Aurojit Panda.

Incremental Specialization of Network Programs

Fabian Ruffy, Zhanghan Wang, Gianni Antichi, Aurojit Panda, Anirudh Sivaraman

Improve Load Balance for DLRM with Programmable Switch

Advisor: Jialin Li (NUS) and Liang Luo (Meta, US)

Deep Learning Recommendation Model(DLRM) is a widely used recommendation model developed by Meta. DLRM uses embedding tables that encode sparse features like movie genres. However, there are a lot of such tables, and the number of embeddings and the embedding sizes vary a lot. Thus, they are usually partitioned into different machines. Nevertheless, the accessing pattern is skewed due to the popularity of data. Thus, we tried to use the programmable switch to improve the load balance by caching several embedding entries and routing the lookup requests based on caching status. The work hasn't been completely done and the problem might already be obsolete.

Database Deadlock Diagnosis for Large-scale ORM-based Web Application

Advisor: Jinyang Li (NYU) and Zhaoguo Wang (SJTU)

Database-backed web application usually relies database to handle deadlocks. However, the common-used detect-and-recover strategy could be costly. Although developers can sometimes reorganize their application to removedeadlocks, the large number of LOC and third-party ORM frameworks they use make this much more difficult. In thisproject, we use symbolic execution to extract the APIs' statement templates with symbolic inputs and path conditions forthe issued statements. Based on the information, we then analyze and report the potential deadlocks. I was only listed in acknowledgement in the paper since I left the team later.