are described in more detail in my research statement.
|
Computer Science Department
Carnegie Mellon University 5000 Forbes Avenue Pittsburgh, PA 15213-3891 |
|
Office: (412) 268-3778
Home : (412) 441-6628 Cell : (412) 657-1353 Fax : (412) 268-5576 rajesh@cs.cmu.edu http://www.cs.cmu.edu/~rajesh/research/ Citizenship: Singaporean, Visa status: F1 student |
Mobile and pervasive computing, operating systems; distributed computing; software engineering; usability of software systems; usability testing; networking
Carnegie Mellon University, Pittsburgh, PA
Ph.D. in Computer Science, 2006 (expected)
Thesis Title: Simplifying Cyber Foraging
Advisor: Mahadev Satyanarayanan
National University of Singapore, Singapore
Masters in Computer Science, July 2000
Thesis Title: CHAMELEON - A System for Adaptive QoS Provisioning
Advisor: Akkihebbal L. Ananda
National University of Singapore, Singapore
B.Sc in Computer Science (Honours), Nov 1998
Thesis Title: Distributed Computing for the CAVE Virtual Reality Display Environment
Advisor: Akkihebbal L. Ananda
My research has encompassed three main areas; 1) mobile computing /
operating systems / distributed systems, 2) software engineering / usability
testing, and 3) network protocols. I present a brief synopsis of this
previous research, beginning with the most recent. Research items marked
with a are described in more detail in my research statement.
Interactive scientific applications typically have two phases; a compute phase that is memory, CPU, or network bound, and an interactive phase where a user interacts with the results of the compute phase. These phases can repeat multiple times. To improve the performance of such applications, I am working with researchers at Carnegie Mellon and the University of Toronto to develop a system, called AgentISR, that uses virtual machines to seamlessly and transparently migrate these applications automatically between a compute farm and the user's desktop machine. This results in the best possible performance during either application phase; the compute phase is able to use the resources available at the compute farm and users can crisply interact with the application locally during the interactive phase. A paper detailing the basic operation of AgentISR has been submitted to EuroSys 2006. For this paper, my key contribution was a user study demonstrating that a thin client solution was ineffective. In addition, I also assisted with the rest of the evaluation. We are currently extending AgentISR by testing its performance with many more applications and by developing policies for automatically migrating applications.
Developer Support for Cyber Foraging (Software Engineering / Usability Testing)
Cyber foraging refers to the opportunistic use of server resources in the environment by a mobile client. These resources allow clients to execute applications that they could not run locally. However, before cyber foraging can become a reality, developers must be able to quickly, easily, and effectively retarget existing applications to support cyber foraging. Previous methods of doing this used to take two to four weeks per application; an unacceptably long time. This problem is made more difficult as a good solution should be usable by even novice developers and should work with applications written in any language and style. As part of my thesis, I developed and rigourously validated a process, called RapidRe, that provides a solution. It allows even novice developers to retarget applications, written in arbitrary languagues, to support cyber foraging in under 4 hours. In addition to dramatically reducing the retargeting time, my work is one of the first that rigourously validates the usability of a software system. This validation was performed via a software usability study involving 13 participants and 8 different applications (written in a variety of languages including C, C++, Java, Tcl/Tk, and Ada). This aspect of my research is described in more detail in my research statement.
Operating System Support for Cyber Foraging (Mobile Computing / Operating Systems)
RapidRe allows developers to quickly and easily retarget applications to support cyber foraging. To ensure that these retargeted applications also had excellent performance, I developed an adaptive dynamic runtime component called Chroma. This runtime is a key component of RapidRe and provides the common runtime components necessary for effectively benefiting from cyber foraging. For Chroma to be effective, I needed to figure out the optimal application settings (fidelity settings and runtime partitioning) that best matched the available resources and user preferences. To do this, I integrated a number of resource measurement modules and a resource prediction module into Chroma. I then built an exhaustive yet efficient solver that picks the appropriate runtime application settings. I also developed an easy and effective way for Chroma to receive user preferences. I then showed that retargeted applications, using Chroma, can achieve excellent performance in a variety of mobile scenarios. A particular novel aspect of Chroma is that it can use extra server resources to automatically improve application performance. I describe this work in more detail in my research statement. The design and validation of Chroma was presented initially at SIGOPS EW 2002 and in more detail at MobiSys 2003.
Distributed Infrastructure for Massively Multiplayer Games
(Distributed Systems)
During my internship at IBM Research Watson, I worked with Maria Ebling, Paul Castro, and Archan Misra to investigate the problem of building a distributed infrastructure for massively multiplayer online games. The scope of the problem was immense as these games could potentially have up to 1,000,000 players in the same game. The key to allowing such a large number of players is to distribute the game environment among many (10,000 or more) different servers; each serving a subset of the total player population. Building and managing such an infrastructure requires overcoming many different challenges. For my research, I concentrated on the problems of a) ensuring that each game server was as consistent as it needed to be, b) developing mechanisms to automatically handle load spikes on any game server (caused by hotspots or other events), and c) ensuring that the infrastructure is attractive to and easy to use by real game developers. We called our infrastructure Matrix. I built a prototype implementation of Matrix and showed that a) Matrix could automatically add and remove servers to handle even very dense load hotspots, b) Matrix has reasonable overheads, and c) Matrix is able to satisfy the demands of real game players (verified with a user study explained below). My research statement contains more details of this work. This work was presented at Middleware 2005 and has been submitted to the IEEE TC Journal.
A key design consideration for Matrix was that it should not add any perceivable latency to a game player's experience. To validate this, I conducted a user study, with the help of Maria Ebling, in the summer of 2004. We asked 5 participants to play a game called Bzflag that had been modified to use Matrix. During the user study, load was introduced and removed from the game to trigger Matrix's load balancing algorithms. Overall, our results were very promising and the participants perceived almost no additional Matrix-induced latency. The full user study details appear in the IEEE TC journal submission.
In this work, I, together with Srinivasan Seshan and Aditya Akella from
Carnegie Mellon University, looked at developing network protocols that
could change behaviour based on the current situation. We looked at two main
problems. First, we looked at ways to modify TCP's behaviour in order to
save power. We showed that by changing TCP's delayed-ack threshold, you
could actually reduce the total amount of energy needed to transfer data, at
the loss of some throughput. Second, we looked at developing an overlay
routing infrastructure that would pick the appropriate overlay routing
algorithm for the current overlay. For example, for small overlays, you
would pick a flat routing algorithm whereas for large overlays, a
hierarchical routing algorithm would be better. This work was presented as a
poster at SIGCOMM 2001 and SOSP 2001.
In Fall 2000, I was asked to port Linux to the iPAQ for a DARPA demo. I used
the ARM Linux distribution and kernel created by Compaq developers as a
starting point. To complete the port, I needed to a) fix bugs in the
infrared network driver, and b) write a new driver to access flash RAM on
the iPAQ. The infrared fixes and new driver were subsequently merged into
the Linux-iPAQ source tree.
As part of my masters thesis, I developed an extension to TCP that improved
its performance over wireless links. The key idea was to add an extra option
to TCP that contained the checksum of the header of the TCP packet. When a
packet got corrupted by the wireless medium, this extra checksum was
checked. If the checksum passed, the information in the header was used to
tell the sender of the packet that corruption of the packet, and not
congestion, had occurred. The sender would then re-send the packet without
reducing its sending rate or timing out. This extension to TCP was
implemented in the Linux kernel 2.2.10 and its performance was tested under
a variety of wireless conditions. The results showed that the extension
significantly improved TCP's performance in the presence of frequent packet
corruption. These results were presented at INFOCOMM 2001.
The other half of my masters thesis involved the use of TCP tunnels to
provide QoS guarantees at gateway routers. TCP tunneling is similar to IP
tunneling except that TCP connections are being tunneled over another TCP
connection instead of over an IP tunnel. The benefits of this approach were
that the tunnels used TCP's congestion control mechanisms and were able to
share bandwidth fairly with other TCP tunnels. I built a system, CHAMELEON,
that used TCP tunnels between gateway routers to provide QoS guarantees for
application traffic using those routers. This work is detailed in my Masters
thesis. I also showed, in a paper presented at LCN 2001, how TCP tunnels
could be used to prevent congestion collapse from occurring at these gateway
routers.
The CAVE is a virtual reality environment, developed by the University of
Illinois at Chicago, that uses a room to completely immerse the user in the
virtual environment. It requires powerful computers to handle the graphics
requirements of the virtual environment. As such, the servers controlling
the CAVE do not have many free CPU cycles to do computation on behalf of
programs wanting to use the CAVE. For my honours thesis, I wrote software
that allowed the computational aspect of an application to be distributed to
other machines, thus leaving the main servers free to concentrate solely on
the graphics requirements of the CAVE. This work was published in a high
performance computing conference
I was an exchange student at the University of California at Berkeley during
the Spring of 1997. While I was there, I worked with Dr. T. T. Blackmon who
was part of the Intelligent Mechanisms Group at NASA Aimes. Dr. Blackmon was
responsible for creating a 3D virtual world from the images returned by the
Pathfinder Probe that landed on Mars in July 1997. On my part, I wrote
software to automatically create 3D meshes from the 2D images returned by
the probe. These meshes were used to create the final 3D world.
| 2003 - 2005 | IBM Ph.D. Fellowship Recipient for 2003/2004 and 2004/2005 |
| This is a highly prestigious fellowship awarded annually to the top Ph.D. candidates worldwide | |
| April 2002 | Agency for Science and Technology Research International Fellowship Recipient |
| This is the Singaporean equivalent of the NSF that awards fellowships to the best Singaporean graduate students | |
| Jan 2002 - Jan 2003 | Awarded a $40,000 personal research grant from USENIX for research on Cyber Foraging |
Professor Mahadev Satyanarayanan
School of Computer Science
Carnegie Mellon University
5000 Forbes Ave
Pittsburgh, PA 15213
Office: (412) 268-3743 Fax: (412) 268-4136
satya@cs.cmu.edu
Professor David Garlan
School of Computer Science
Carnegie Mellon University
5000 Forbes Ave
Pittsburgh, PA 15213
Office: (412) 268-5056 Fax: (412)-268-3455
garlan@cs.cmu.edu
Professor Srinivasan Seshan
School of Computer Science
Carnegie Mellon University
5000 Forbes Ave
Pittsburgh, PA 15213
Office: (412) 268-8734 Fax: (412) 268-5576
srini@cmu.edu
Doctor Maria R. Ebling
Research Staff Member
IBM T. J. Watson Research Center
P. O. Box 704
Yorktown Heights, NY 10598-0704
Office: (914) 784-7949 Fax: (914) 784-7455
ebling@us.ibm.com
Professor James D. Herbsleb
School of Computer Science
Carnegie Mellon University
5000 Forbes Ave
Pittsburgh, PA 15213
Office: (412) 268-8933 Fax: (412) 268-7287
jdh@cs.cmu.edu
Additional references available on request.