Contact email: abanerjee at ucdavis dot edu
I am currently at the performance engineering team in VMware. I have worked with performance and scalability challanges at Oracle, Sun Microsystems, and as a Ph.D. researcher at UC Davis in the past. This webpage has details of my research work associated with UC Davis. For my current activities, please visit my blog, and my Linked In profile.
My paper titled "A Review of Wavelength-Division Multiplexed Passive Optical Network (WDM-PON) Technologies for Broadband Access" in JON, October 2005, is prominently cited in WDM-PON literature.
Our paper on "Introspective End-system Modeling to Optimize the Transfer Time of Rate Based Protocols" has been accepted in ACM HPDC, 2011. Paper and slides coming soon.
"Active End-System Analysis to Estimate the Network I/O Bottleneck", presented at IOM-2009, Raleigh, NC. Presentation available here.
"Algorithms for Integrated Routing and Scheduling for Aggregating Data from Distributed Resources on a Lambda-Grid," to appear in IEEE Transactions on Parallel and Distributed Systems, Vol. 18, No. 1, January 2008. paper
"Multi-Thread Polling: A Dynamic Bandwidth Distribution Scheme in Long-Reach PON," to appear in IEEE Globecomm, November 2007.
"Improving Data Transport over High-Speed Networks," PhD dissertation, UC Davis. dissertation.
"Modeling and Analysis to Estimate the End-System Performance Bottleneck Rate for High-Speed Data Transfer", Proc., PFLDNet 2007, Marina Del Rey, CA. paper, slides.
"A Protocol for Efficient Tunable Laser Utilization to Support Incremental Upgrade in a WDM-PON" has been accepted for presentation at OFC 2007.
"Fair Sharing Using Dual Service-Level Agreements to Achieve Open Access in an Ethernet Passive Optical Network (EPON)", In IEEE Journal on Selected Areas in Communications, August 2006. paper.
Research InterestsModeling and Analysis to Estimate the End-System Performance Bottleneck Rate in High-Speed Networks
The maximal rate at which network Input/Output (I/O) may be performed may exceed the end-machine(edge) processing capacity to handle it. As an example, when the network line rate is 10 Gbps, the end-machine may not be able to handle with data arriving given its current workload and therefore it may often be a bottleneck to data transfer. This performance bottleneck is dependent on the workload at the system. A key deficiency in most transport protocols is lack of an accurate determination of such a bottleneck rate at which the receiving system can handle network I/O. In this work we propose an analytical model of a computing system to determine the optimal rate at which the best network performance may be extracted from the same. This optimal rate is delivered to the sender to achieve better transport performance. We have developed a transport protocol which integrates this feedback for adjusting the sending rate. This work is in submission to IEEE Transactions in Networking (TON).
End-system Performance Feedback for High-speed data transfers (RAPID).
As optical networks penetrate more towards end-machines (e.g., home desktops via the PON, scientific computing machines via the lambda grid), the network speeds may outperform the end-system processing capabilities. Thus congestion may be shifted from the network to end-host machines which may not be able to keep up with the high-speed data rates at all times. "Rate-Adaptive Protocol for Intelligent Delivery (RAPID)" is our proposal for a lightweight end-system performance-aware transport protocol. Based on the self-monitoring of the dynamic task-priority at the receiving end-system, RAPID enables the receiver to proactively deliver feedback to the sender to adapt its sending rate to avoid congestion at the receiving end-system. This avoids the large bursts of packet losses typically observed in current transport protocols. Over a 10-Gigabit link emulation of an optical circuit, RAPID reduces file-transfer time by as much as 25% compared to the Reliable Blast UDP (RBUDP) protocol.
Scheduling file transfers over Lambda Grids
A Lambda Grid is a circuit-switched optical backbone network which allows point-to-point dedicated connections to be set up. In many distributed computing applications, dedicated circuits must be established for large-scale data transfers, e.g., from data warehouses to a processing node. My research investigates the on-line and off-line scheduling algorithms for reservation of circuits on a lambda grid. Given a set of files, the file sizes, the locations of the files on the lambda grid, and a common destination (such as a supercomputer located on the network), as inputs, the algorithms perform intelligent routing (in space dimension) and scheduling (in time dimension) of lambda grid circuits for fast data aggregation.
Open Access for services in a Passive Optical Network (PON)
The PON is an optical-fiber-based technology for residential subscriber access networks. In the open-access framework, an access-network channel is shared by multiple competing service providers to deliver services to independent subscribers. My research investigates the use of two degrees of Service-Level Agreements (called Dual SLAs) to meet fairness in terms of throughput and network latency in such a shared access-network channel. Fairness is met among users located at one end of the channel, and competing service providers at the other end.