One important area of systems research deals with the interface between computer hardware and software. Solid research in this area--often called "computer architecture"--must penetrate beyond an interface to deal with issues in one of more the following areas:

• hardware implementation (e.g., exploring how trends are leading wire delays to dominating gate delays, thereby making global bypassing more expensive),
• system software (e.g., translation lookaside buffer alternatives to support large page sizes),
• compilers and instruction set architecture (e.g., scheduling loads together for non-blocking caches),
• applications (e.g., large buffer pools for databases or reserved memory bus bandwidth for multimedia), and
• sophisticated performance analysis methods (e.g., complex simulators, mapping hardware performance counters back to application measures, analytic models).

In my view, several important areas of research for the next 5-10 years are:

• Memory systems. Memory system behavior will continue to determine SUSTAINED performance, as the gap between instructions per second and memory latency continues to grow dramatically. Surprisingly, a load today is far more complex than a floating-point divide. A load can cause an effective address calculation, TLB access, L1 cache access (and replacement), L2 cache access (and replacement), and memory access. In shared memory multiprocessors, considerable coherence protocol activity is also possible. More work is needed to make memory systems more stable, especially in the presence of audio and video streams. Ken Kennedy's keynote at PPoPP97 predicted ten levels of memory hierarchy in large machines!
• Speculation everywhere. There are fundamental limits to what software can know statically. Recent work on dynamic speculation has moved way beyond branch prediction to path and value prediction. Value prediction opens up the possibility of executing a program with k operations on the critical path in less than k cycles.
• Support for new datatypes and applications. While the past belongs to floating-point numbers and character strings, the future belongs to graphics data aggregates (e.g, MMX (TM)), JPEG, MPEG, etc. It's not your father's datatype any more! Furthermore, more support will be needed for (de-)compression and (de-)encryption, the cornerstone of authenticated, private communication in science and in life.
• Network-centric computing. More and more computers will facilitate communication and only incidently compute (to transform the data streams being communicated). Thus, networks, network interfaces, and methods of virtualizing them without loss of performance or protection will become more important. Someday two people should be able to have a video conference with no operating system touching data after the initial set up.
• Parallel computing. Parallel computers are entering the mainstream with Quad Pentium Pro (TM) daughter boards. Nevertheless, many the our nation's needs will require greater power than the market place and Moore's law delivers. While parallel computing already has had some spectacular successes, performance is not yet robustly good.
• High Availability. During my sabbatical in industry I grew to appreciate the need to make computing more robust (since we are building much of our nation's infrastructure on it), but, at the same time, the unwillingness of customers pay for serious fault tolerance (a la deep space probes). The compromise that needs to work is high availability, which allows some components to crash, but keeps other components running and re-incorporating recently-crashed components quickly. An example of what we need is a network-of-workstations web server, which will occasionally drop connections, but will almost always be up, and, at the same time, not pay any overheads when all components are up.

Postscript

I would strongly advocate that academics, especially in systems research take sabbaticals in industrial product groups. For more information on this position and my experience, please see URL "http://www.cs.wisc.edu/~markhill/nsf_es96_pos.ps".