Computer Science Thesis Proposal

Monday, May 15, 2017 - 11:00am


Traffic21 Classroom 6501 Gates Hillman Centers


JIN KYU KIM, Ph.D. Student

Machine learning (ML) methods are used to analyze data which are collected from various sources. As the problem size grows, we turn to distributed parallel computation to complete ML training in a reasonable amount of time. However, naive parallelization of ML algorithms often hurts the effectiveness of parameter updates due to the dependency structure among model parameters, and a subset of model parameters often bottlenecks the completion of ML algorithms due to the uneven convergence rate. In this proposal, I propose two efforts: 1) STRADS that improves the training speed in an order of magnitude and 2) STRADS-AP that makes parallel ML programming easier. In STRADS, I will first present scheduled model-parallel approach with two specific scheduling schemes: 1) model parameter dependency checking to avoid updating dependent parameters concurrently; 2) parameter prioritization to give more update chances to the parameters far from its convergence point. To efficiently run the scheduled model-parallel in a distributed system, I implement a prototype framework called STRADS. STRADS improves the parameter update throughput by pipelining iterations and overlapping update computations with network communication for parameter synchronization. With ML scheduling and system optimizations, STRADS improves the ML training time by an order of magnitude. However, these performance gains are at the cost of extra programming burden when writing ML schedules. In STRADS-AP, I will present a high-level programming library and a system infrastructure that automates ML scheduling. The STRADS-AP library consist of three programming constructs: 1) a set of distributed data structures (DDS); 2) a set of functional style operators; and 3) an imperative style loop operator. Once an ML programmer writes an ML program using STRADS-AP library APIs, the STRADS-AP runtime automatically parallelizes the user program over a cluster ensuring data consistency. Thesis Committee: Garth A. Gibson (Co-chair) Eric P. Xing (Co-chair) Phillip B. Gibbons Joseph E. Gonzalez (University of California, Berkeley)Copy of  Proposal Summary

For More Information, Contact:


Thesis Proposal