Professional Narrative

I received my Ph.D. at Berkeley under Prof. Eugene Wong in 1980, working on database design and translation within the Ingres Relational Database project. I had the good fortune to have also worked one summer with the System-R Relational Database Group at IBM's San Jose Research Laboratory, so I was located in the center of the relational database design and implementation universe.

I spent one year working at BBN, working with the legendary Dave Walden and John McQuillan (inventors of the original Internet routing algorithms), and at Computer Corporation of America in Cambridge, MA, a leading database research center at that time. In 1981, I moved to the University of Wisconsin-Madison, where I shifted my interests into VLSI and CAD, in particular, database support for VLSI CAD environments. An opportunity presented itself to return to Berkeley in 1983, which I accepted gladly, selling my snow tires. Since that time, my research and teaching interests have focused on a series of projects spanning the design, engineering, and implementation of a variety of advanced computing systems spanning hardware and software.

Research Projects at Berkeley

• SPUR Project: In 1984 I switched from CAD to processor and memory system design. Between 1985 and 1989, I led the design of the distributed cache and virtual memory organization of the SPUR multiprocessor project. We invented the term "snooping caches" (also known as "snoopy" caches), and implemented one of the first invalidation protocols. The overall SPUR project was led by Dave Patterson, and involved Richard Fateman, Paul Hilfinger, Dave Hodges, and John Ousterhout. Snooping cache schemes are used in many multiprocessor servers on the market today.



• RAID Project: In 1986, I had an idea for building high performance storage systems by ganging together a large number of small formfactor disk drives, much like the processors in a multiprocessor. Between 1987 and 1992, I led the design and implementation of the Berkeley RAID (Redundant Arrays of Inexpensive Disks) high performance storage system (we coined the term "RAID"). This work was done in overall collaboration with Dave Patterson and John Ousterhout. Click here for a picture of "RAID the First" and here for a picture of "RAID the Second," which is now on exhibit at the Computer History Museum in Mountain View, California. This is the innovative "shelves of disks" that gave us such incredible MBytes per square foot for its time (72 3.5" disks in a 19" rack). This multibillion dollar per year industry all started with a simple idea in Berkeley in 1986! Too bad I didn't get any royalities.



• Wiring the White House: After the elections of 1992, I took a leave during 1993-1994 at the Computing Systems Technology Office (now "Information Technology Office") of the Defense Advanced Research Projects Agency. I started as a program manager and ended as an office deputy director. Along the way, I was responsible for establishing whitehouse.gov and the president and vice.president Internet mail accounts, I participated on Vice President Gore's National Performance Review, and played a lead role in formulating the Federal HPCC program's Information Infrastructure Technology and Applications research program in support of the Administration's National Information Infrastructure Initiative.
  To see a photograph of me with President Clinton and Vice President Gore, click here. To see the letter written to me by Vice President Gore, click here for plain text, and here for a scan of the letter.
  To read about my Washington experiences, see Prof. Katz WENT to Washington.



• Bay Area Research Wireless Access Network: When I came back from Washington in 1995, I began working on Mobile Computing and Wireless Communications. With Prof. Eric Brewer, we focused on issues of wireless/wireline network integration, internetwork operation, and end-to-end quality of service guarantees in the context of a local-area wireless testbed in Soda Hall and a wide-area testbed located in the San Francisco Bay Area. This effort was called the Daedalus Project. The combined in-building, metropolitan-area, and wide-area testbed was called BARWAN (Bay Area Research Wireless Access Network). This work was supported by the Defense Advanced Research Projects Agency under its GloMo Program, as well as numerous industrial sponsors. Major developments included low latency handoffs for real-time streams, intelligent vertical handoffs between overlay networks, improved transport layer performance over wireless and asymmetric links, and proxies for intelligent data adaptation for web data types and real-time streams. This project reached its successful conclusion in August 1998.



• MASH: This was led by Steve McCanne and myself between 1997 and 2000, once again under DARPA sponsorship, to scale up the audio, video, and whiteboard tools that Steve had developed as part of his dissertation. My interest focused on video archiving and building subscription-based applications within the multicast framework developed by Steve. Prof. Larry Rowe subsequently created the Open Mash organization to further develop the software as part of his interest in webcasting.



• NINJA: At the same time that MASH was going on, I worked with Eric Brewer, David Culler, and Anthony Joseph on a new network service architecture called NINJA. As a cluster-based scalable execution and service environment, it unified Culler's work on server farms (NOW), with Brewer and Joseph's interests in distributed systems, with my own interest in network support for services. See Ninja.
  
  
• ICEBERG: Following the success of BARWAN, Anthony Joseph and I began a close collaboration with Ericsson to investigate the architecture of "Beyond Third Generation Telecommunications Systems." ICEBERG developed a wide-area service creation and management environment for computer-telephony integration, merging ideas from NINJA and BARWAN. It ran from 1998 to 2001, unfortunately coinciding with the Internet and telecommunications boom and bust.



• SAHARA: The impact of ICEBERG was limited in that it required universal deployment of the architecture to effectively support new kinds of telecomms services. This observation was the seed for SAHARA, an "interoperability" architecture that formed a new kind of service glue to tie different kinds of edge networks together. The project received support from several telecomms equipment companies and service providers between 2001 and 2004.



• ENDEAVOUR: Overlapped with ICEBERG and SAHARA was Endeavour, a broad research investigation into future computing architecture based on harnessing the extremes spanning from MEMS to large-scale cluster machines distributed across the world. The project was one of the four DARPA "Expeditions" contracts. The common theme across the diverse projects within Endeavour was the development of a reliable, scalable, composible, wide-area service architecture to support diverse access networks, end devices, and services that can operate on a global-scale. Endeavour involved over a dozen Berkeley faculty and three dozen graduate students, and laid the foundation for many follow-on projects, including OceanStore, Telegraph, and SensorNets.



• OASIS: I have recently gotten interested in developing network elements that can inspect and act upon flows inside the network, particularly to protect the network in the face of intrusions, failures, or overload. Routers are becoming ever more capable in terms of deep packet classification, even at gigabit speeds, yet little is understood about how to make use of this ability to comprehensively observe and manage the network. OASIS started in 2004 as the logical follow on of SAHARA.



• RADLab: RADLab is newly established industry funded project (Google, Microsoft, and Sun) investigating the architectures needed to develop, deploy, and manage the next class of Internet-scale applications. It is a collaboration between Armando Fox, Mike Jordan, Dave Patterson, Scott Shenker, Ion Stoica, and myself. Statistical learning, how to implement it and exploit it in a systems context, is a major theme of this project. We plan to use statistical learning approaches, based on observations at the network layer, with extensive cross-layer interactions, to form the basis of our approach to pervasive network management and protection.