About

I am Sanket Purandare, 3rd year PhD student at the Data and Systems Lab (DASLab), School of Engineering and Applied Sciences (SEAS), Harvard University. I am advised by Prof. Stratos Idreos. My broad research interests lie in the area of systems for deep learning. In particular, I work on the characterization of deep learning training and inference workloads and optimizing the deep learning software infrastructure for hardware efficiency by applying systems concepts from operating systems, compilers, and computer architecture in a novel fashion. 

Currently, I am working as a student researcher with Meta Inc. (formerly Facebook). I work with the PyTorch Compilers and Distributed Teams on buidling the compiler stack for PyTorch distributed training. In particular, my work involves doing compiler level graph optimizations for distributed training algorithms in order to minimize the communication overhead and enhance the scalability of the approaches. Prior to that I worked as a Research Scientist Intern with the same team in Summer 2022 under the mentorship of Dr. Animesh Jain.

Before joining Harvard, I spent wonderful six months as a Research Intern at Microsoft Research India (MSRI) under the mentorship of Dr. Karthik Ramachandra on the project Aggify. We designed and implemented an algorithm for semantically mapping inefficient imperative PL-SQL code to SQL and hence making it amenable to rich query optimization framework inside Microsoft SQL Server.

Before that, I was a Research Associate at the Database Systems Lab under the guidance of Prof. Jayant Haritsa after graduating with a masters from the Indian Institute of Science (IISc). In my first project, we worked on a family of algorithms that perform query evaluation without estimating the selectivity of the predicates in the query. My thesis theoretically lowers the worst-case performance bound of these algorithms and verifies the same empirically in PostgreSQL. In the second project, we worked on a machine learning-based taxonomy system that decides whether one should use the native optimizer’s query plan or some other robust query processing algorithm by predicting a risk score corresponding to the cardinality estimates. Both projects were co-mentored by Dr. Vinayaka Pandit, IBM India Research Labs.

Please see my resume for more details.

Outside academia, I am a Fellow at the Harvard GSAS Student Center actively involved in building a student community at Harvard that promotes diversity, inclusion, and belonging by hosting a range of events that cater to a wide variety of student interests and cultures. I like to travel and explore new cities and cultures. I am a motorcycle and car enthusiast and an amateur painter.

Selected Publications

Aggify: Lifting the Curse of Cursor Loops using Custom Aggregates.
Surabhi Gupta, Sanket Purandare, Karthik Ramachandra.
ACM International Conference on Management of Data (SIGMOD), 2020.

Dimensionality Reduction Techniques for Bouquet-based Approaches
Sanket Purandare
Masters Thesis, Indian Institute of Science