[CS] Announcement of Cesar Andres Stuardo Candidacy Exam

[nitayack at uchicago.edu]

This is an announcement of Cesar Andres Stuardo's Candidacy Exam.
===============================================
Date: August 5, 2021

Time: 5:00 pm

Location:  Join Zoom Meeting
https://uchicago.zoom.us/j/95754515100?pwd=WU01enpsbmE1eiszMW9WRVlKQ2I0dz09

Meeting ID: 957 5451 5100
Passcode: 294475

Candidate: Cesar Andres Stuardo

Title:  Towards Automated Scalability Fault Detection and Behavior Modeling on Modern Large-Scale Cloud Systems.

Abstract: In light of the limits of Moore’s Law, Dennard scaling and the ever-increasing computing demand, the last decade has seen unprecedented deployment scales. There are undeniable benefits on the rise of large-scale cloud systems, but is scale a friend or a foe? On the positive side, scale surpasses the limits of a single machine in meeting increasing demands in compute and storage. On the negative side, scalability is not an after-thought: It requires systems to be designed with it in mind, as many of their protocols could seem scalable when evaluated using small inputs (e.g.small amounts of data) or under low load but explode in terms of cost (e.g.execution time, memory consumption) when the size of such inputs or load grows. We refer to these types of explosive manifestations as scalability faults, i.e.faults that become visible at larger scales but are not visible at smaller scales.

In our research, we conduct a study that involves more than 400 scalability faults in over 14 open source cloud systems. We analyze these faults and identify the related root causes, categorize their structural code patterns and analyze the available mechanisms used to find them if any. Based on our observations, we propose the creation of an automated scalability fault detection methodology, comprised of a pipeline that combines new static analysis and dynamic profiling tools with existing test mechanisms and frameworks to automatically find such scale sensitive code paths. Furthermore, we propose the creation of ML-based mechanisms to model the behavior of such code paths as a function of their scale dependencies, allowing for accurate extrapolation of the target metrics (e.g.execution time, memory consumption) over larger and previously unobserved scales.
Advisors: Haryadi Gunawi

Committee Members: Haryadi Gunawi, Shan Lu, and Jeffrey Lukman