Vijay Vasudevan

IllustratedSketch
Where I work
Graduate Student
Computer Science Dept.
Carnegie Mellon University
WeH8101
email: click or make an educated guess
I am a Ph.D candidate in the Computer Science Department at Carnegie Mellon University. I am co-advised by Professor David Andersen and Professor Hui Zhang.

Research Projects

My interests span topics in networked computer systems and Internet architecture. In particular, I am motivated by discovering solutions that are practical, deployable, and flexible.

Simple Wide-Area Multi-Path (SWAMP)

We are exploring routing architectures to improve Internet availability using multiple concurrent interdomain end-to-end paths. Our first design focuses on a simple, effective, and economically-friendly architecture to promote real-world deployment. We find that providing path choice at both a multi-homed source and destination can provide nearly as many AS disjoint paths as a policy-free source routing mechanism could. Please see our technical report for measurement details and analysis. Based on this result, SWAMP can provide end-hosts with several disjoint paths in an architecture conducive for deployment. More details on SWAMP to come.

Collaborators: David Andersen, Hui Zhang

Incast: TCP Throughput Collapse in Cluster-based Storage Systems

In TCP-based high-bandwidth, low-latency storage area networks where one client synchronously requests data striped across many servers, TCP throughput has been observed to be one or two orders of magnitude less than the client's link bandwidth. This throughput collapse poses a problem for storage servers running on commodity hardware (Ethernet/IP), especially considering the recent push for a 100GigE Ethernet standard for high-performance networked computing. We have successfully reproduced the Incast phenomenon in simulation and real-world deployments and are currently investigating solutions to remedy this collapse.

Collaborators: Elie Krevat, Amar Phanishayee, David Andersen, Greg Ganger, Garth Gibson, Srini Seshan.

FAWN: Fast Array of Wimpy Nodes

In today's typical system architectures, high-speed, power-hungry multi-core processors are separated from persistent storage devices via a hierarchy of fast, volatile memory. However, for workloads which do not benefit from locality of reference, such as query-driven workloads (e.g. DNS), the benefit of a high-speed processor is wasted. This increasing CPU-memory-disk gap actually creates an exciting opportunity, which is fueled by three observations: 1) The throughput and seek times of modern magnetic disks have not kept pace with the increase in CPU speeds. 2) Power consumption increases super-linearly with speeds. 3) Flash memory is now cheaper per gigabyte than magnetic disk for 4GB or less. By placing cheap, flash storage next to small energy efficient processors, we restore the balance between processing power and data storage capabilities to obtain a scalable, energy-efficient, and effective server architecture for an important class of query-driven and scan intensive workloads.

Collaborators: Jason Franklin, Amar Phanishayee, David Andersen Publications

Technical Reports

Presentations

Past Projects

As an undergraduate at Berkeley, I worked on two projects in Ken Goldberg's Berkeley Automation Lab. Specifically, I was the network engineer on the Ballet Mori project and also worked on physical deformation simulations for needle surgery planning. I also worked on DDoS prevention in next generation Internet architectures and implemented ConeNAT functionality within the OCALA framework.

Friends

CMU

David Brumley
Jason Ganetsky
Swapnil Patil
Amar Phanishayee
Kanat Tangwongsan
Dan Wendlandt

Elsewhere

Beverly Wang
Kevin Lin
Chris Crutchfield
Robert D. Gregg