My research lies broadly in the areas of cloud computing, virtualization, data center management and computer architecture with an emphasis on resource management and power-efficient/green computing systems. My research interests span data center energy management, virtualization resource and power management, computer architecture and its interaction with systems software; architectural and real-system techniques for power-efficient, reliable and secure computing systems; hardware/software performance monitoring; microprocessor power/thermal modeling and real measurement techniques; phase analysis; live, runtime monitoring and prediction of workload phase behavior and phase-driven power and thermal management.

Origami: Systems as Data: Recently [2012-2013], I am working in the the Scalable Data Center Analytics team in IBM Research, on a very interesting project called Origami. The main idea of this project is treating systems as data and "touchless" monitoring and management of virtualized systems in the cloud. My main goal is to bring touchless monitoring into running VMs, without agents, side effects or any visible impact on the running system. We have two great collaborations with Prof. Mahadev Satyanarayanan, and his group from CMU, and Prof. Eyal de Lara and his group from UofT. As part of these collaborations, I have the privelege of working with Wolfgang Richter (CMU) and Sahil Suneja (UofT).

• A first incarnation of the Origami project is in IBM Virtual Image Library(demo video).
• Wolf's work on gamma-ray is open sourced and the qemu patch may be included in KVM distro (github).
• Nice Origami overview by Vas on Research Blog.
• Origami ASPLOS Tutorial: System Analytics in the Cloud.

Thomas: The Data Center Robot: Since 2010, I have been working on a Robotic Data Center Monitoring and Management project. We started the project with Jon Lenchner and Chris Mansley as a way of autonomously mapping and monitoring data centers, and more recently we have been looking into using a similar solution for daata center asset management--a problem with almost no research on, while a major pain point for data center managers. We received a best paper, best of show, and other awards for the data center robot, it has been deployed on 4 continents in 9 data centers so far, and it has been an immensely fun project.

• A collection of videos for the robot (check out the bloopers as well:)): Thomas Videos.
• Jon's Maps page for tracking deployed robots: Data center robots around the world.
• STG Lab Services robotic measurement offering: Document and Pat's description video.
• External references:
  - Slashdot: IBM Uses Roomba Robots To Plot Data Center Heat
  - Data Center Journal: IBM’s Energy Saving Robot
  - Seeking Alpha: iRobot Will See Future Revenue Boost From Brazil Contract And IBM Technology (This was fun:), and also fell into my iPhone stocks app for IRBT, will add snap here soon.
  - more external refs...
  

S3Virt: Agile Virtualization Power Management with Low-latency Server Power States: I worked with Jeff Kephart, Suzanne McIntosh, Jim Hanson, Rajarshi Das from IBM Research, and many other colleagues across many divisions of IBM on this project. I also led three joint programs over three years related to this work. It was one of the most excrutiating, yet fun;) projects I probably worked on, involving work around server hardware, firmware, OS, virtualization and middleware. I wish I had a picture of Sue sitting in the cold aisle with her coat on, with the octopus cable, and a video of us pulling the blades switching the special switch, and reinserting boards after each experiment on the early stages. I will write more about this at some point, as we finally were able to write a paper about it.

• Video of autonomous, S3-based virtualization power management:S3Virt demo.
• Video of one of prototype systems sleeping:) (Sleepy server video).

VM Demand Estimation: One of my first projects at IBM, working with Gosia Steinder and the Vespa team. The idea behind this was, "it is easy to squeeze, but hard to expand", in the virtualization power management context:) (kinda rhymes). We developed a novel demand estimation technique that tells what a VM wants, even it is under heavy resource contention. This was included in IBM VMControl (KVM Demand Estimation) and Pure System Application Server (VMW demand estimation).

• VMControl even has a nice GUI of Estimated Demand:), here is a screenshot, and a video showing estimated demand in action.
• IPAS related documentation ([tba]).

VM Migration Characterization and Improvement: The main driver of this project was undeerstanding the limitations of VM migration and develop techniques for improving VM mobility and optimizing link utilization. We showed that via various optimizations, we could achieve a quite significant (about 50X) improvement, from our initial implementation, which was relatively a strawman, generic implementation. The more interesting bit was when we involved RDMA, instead of standard TCP migration, the overheads and latency fell off a cliff, which was very interesting to see.

• The RDMA-based migration implementation is now included in KVM, thanks to Michael Hines: QEMU feature documentation.

TODO: I have not yet linked papers/patents to their corresponding projects. Please see the publications page for the related papers, which also have a more comprehensive account of all my collaborators on various projects.

The links below point to additional current and previous research pages:

My IBM Research Page

My Princeton Research Summary (Long Version)

My Princeton Research Summary (Short Version)