Monday, February 11
12:30-1:20pm
JCL 390 [note new location]
Scaling Machine Learning and Analytics with Understanding Isolation
This is an announcement for the Pizza Seminar, a grad student initiative where PhD students (and sometimes postdocs!) present to each other over a free lunch. Talks can be a chance to present something from your research of general interest, practice a conference presentation, or just tell us about something interesting.
The talk will take place in JCL 390 on Monday, February 11. Zechao Shang will be giving a practice job application talk; the abstract and speaker bio are attached,
Scaling Machine Learning and Analytics with Understanding Isolation
Abstract: The increasing demand for application performance require systems to scale massively. Strong isolation, or the guarantee that concurrent operations to behave as they are serial, is critical for applications to behave correctly in concurrent environments, but is also a major scalability bottleneck. Weak isolation sacrifices correctness for performance and often causes isolation anomalies that result in problematic behaviors.
This talk presents two projects that identify and resolve scalability challenges in different isolation contexts. First, I present a new isolation-anomaly monitor, RushMon, for large-scale optimization algorithms, such as stochastic gradient descent (SGD), that are tolerant of isolation anomalies. We identify whether weak isolation compromises the optimization’s convergence. The key observation is that excessive isolation anomalies affect the quality of concurrent operations. RushMon extends the database transaction processing theory to the application of large-scale optimization, monitors the isolation anomalies in real time, and alerts the end-users when the convergence is in question. Second, for isolation-anomaly-aversive applications, I propose how to accelerate the component for enforcing strong isolation. Specifically, I present a novel hybrid concurrency control protocol that utilizes the vertex degree distribution of the graph to help graph analytics scale out.
Bio: Zechao Shang is a postdoctoral scholar at Department of Computer Science at the University of Chicago working with Prof. Michael J. Franklin and Prof. Aaron J. Elmore. His research interest centers on the design and implementation of large-scale data management systems, especially those with applications on machine and data mining. He is on the academic job market and this is a preview of his presentation.