MIT EECS - eBay Undergraduate Research and Innovation Scholar
Emulating Datacenter Network Protocols with Fastpass
Data centers, with their unique mix of flows and application demands, are constantly facing challenges due to heavy queueing. Often, the flows can be highly latency sensitive and thus, minimizing queueing is a priority. Fastpass is a data center network architecture where all senders delegate control over packet transmission to a centralized arbiter. This arbiter selects the path to send traffic over as well as the timeslot to send traffic for each packet, to create a highly efficient no-queueing network. Our goal is now to implement different scheduling algorithms that select timeslots and paths, and evaluate their performance with realistic workloads on real datacenter infrastructure. We will start with known algorithms, and as we gain insight we will modify these algorithms or develop new ones.
This summer I worked with the Analytics team at Pandora, developing a framework to facilitate detection of inefficient queries run on their Hadoop cluster. Additionally, the previous summer I developed an algorithm and accompanying software to predict customers’ paths through grocery stores during an internship with Catalina Mobile.