Research Interests


I am interested in computer architecture with emphasis on microarchitectures of general-purpose, high-performance microprocessors -- superscalars, simultaneously-multithreaded (SMT) processors and chip multiprocessors (CMP). I investigate techniques to achieve high performance, low power, and fault tolerance. I am also interested in hardware for networking such as high-speed routers. My projects and students are listed here and please follow this link to my publications.

Addressing hard errors (or defects), my PhD student Ethan Schuchman and I propose a novel defect-tolerant and testable microarchitecture.

In our work on wire delays, my PhD student Zeshan Chishti and I have argued that SMTs do not have a wire-delay problem . To manage wire delays in large caches, Chishti, Powell and I propose NuRAPID . Recently, we have proposed a new CMP cache to alleviate wire delays in CMPs.

In the soft-error project, Prof. Irith Pomeranz, my PhD students Mohamed Gomaa and Chad Scarbrough and I propose using SMT and CMP to provide recovery for SMT and CMP without much performance loss beyond detection. Recently, Gomaa and I have developed techniques to trade-off coverage for performance.

In the low power project, Profs. Kaushik Roy and Babak Falsafi, my PhD student Michael Powell, and I did one of the earliest work on leakage power reduction in caches (which are the biggest problem for leakage)-- gated-vdd , DRI cache . Addressing power-related issues, Powell and I have proposed schemes damping and tuning to reduce inductive noise . In a recent paper, Powell, Gomaa and I propose exploiting CMPs and SMTs to alleviate power density.

In the speculative multithreading project, Prof. Rudi Eigenmann, his PhD student Troy Johnson, and I proposed a min-cut based technique to partition sequential programs into speculative threads.

In the networking hardware project, my PhD student Jahangir Hasan, my colleague Satish Chandra, and I proposed techniques for efficient use of DRAM bandwidth to improve network processor packet throughput . More recently, Hasan and I proposed an IP Lookup scheme that scales well in all the five factors of total memory, throughput, power, update cost, and chip area.