|
|
Research Interests My research interests are in the areas of networking and operating systems. My current work has been motivated by the fact that heavyweight, kernel-resident network protocols are often unable to take advantage of high-speed network fabrics such as ATM and Fast Ethernet. In order to address this issue, I have developed a framework for user-level network protocols. User-level network protocols are attractive because they can be tailored for specific applications, thus avoiding the performance overheads associated with all-purpose resource management policies in the kernel. My work has focussed on operating system support for user-level access to network hardware and language-cum-compiler support for efficient protocol construction. In the operating systems area, I have worked on the design and implementation of U-Net, a user-level communication architecture that allows user-level processes to access the network hardware directly without violating protection. By removing the operating system from the performance-critical send and receive paths, U-Net is able to demonstrate an order of magnitude improvement in latencies and bandwidths. We have implemented both legacy protocols like TCP/IP and newer protocols like Active Messages over U-Net. The low message-latencies over U-Net also make it possible to run fine-grained parallel programs over workstation clusters. For example, we have shown that Split-C programs running on a cluster of Sparc20s connected by a 140 Mbit ATM network and using U-Net have performance comparable to that on MPPs such as the CM-5, the Meiko CS-2 and the SP-2. In the languages and compiler area, I have worked on developing Promela++, a language that facilitates construction of correct and efficient user-level network protocols. Promela++ is an extension of Promela, a protocol validation language developed by Gerard Holzmann at Bell Laboratories. Promela++ allows automatic verification of program correctness by checking for violations of programmer-specified safety properties. In addition, the Promela++ compiler generates efficient C code from protocol specifications using a variety of domain-specific optimizations. We have implemented the Active Messages protocol in Promela++ and checked some elementary safety properties using the Promela protocol validator. The compiled C version shows performance on-par with the hand-coded version. We are currently in the process of implementing more aggressive optimizations for the Promela++ compiler and implementing several layers of the Ensemble group communication system using Promela++. I have also worked on theoretical aspects of fault-tolerance. We have been able to characterize a class of problem specifications that can be solved in the presence of processor crashes and message losses. We have also come up with a transformation that automatically converts any algorithm that solves the problem when the links are lossy too. |