Low-performance Networking and the Mildly-Parallel Palmtop Array (MPPA) Subcomputer
Joe Futrelle
National Center for Subcomputing Applications, University of Illinois
Urbana-Champaign, Illinois
futrelle@ncsa.uiuc.edu
In this paper, we describe the application of new low-speed networking technologies to our flagship subcomputing platform, the Mildly-Parallel Palmtop Array. In particular, recent research demonstrating that light can be slowed and even stopped will allow us to approach previously-unattainable nanobaud transfer rates in our fiber-optic network.
The scientific enterprise is increasingly burdened with trivial problems, such as which Powerpoint auto-content template to use for the next conference paper. Until recently, problems this miniscule were solved with vastly overpowered computing hardware such as desktop PC's, wasting countless CPU cycles which could have been pressed into service for such worthwhile endeavors as the Search for Extraterrestrial Intelligence[1]. The solution to this burgeoning problem is to provide scientists with access to cutting-edge subcomputing capability.
Our architectural strategy for subcomputing is simple, but revolutionary: use small numbers of slow processors connected with high-latency, low-volume, unreliable networking. We've selected the Palm platform as our main processing unit, because we have been able to achieve extremely high price/performance ratios by selecting models with snazzy titanium cases[2]. The Palm processors are mounted in a rack and networked together using infrared beams. Small mirrors are used to direct the beams from one processor to the next. By neglecting to clean the mirrors, we were able to reduce the transfer rate considerably.
Codes developed for the Mildly-Parallel Palmtop Array are written using special parallel programming language called Fermata, which distributes elements of the computation across the couple-three CPU's in the array[3]. For instance, on a conventional desktop workstation, the computation "1+1" is carried out with a single instruction. In contrast, to compute this on the MPPA, one processor handles the first "1", a second manages the "+", and a third takes care of the second "1". Completing the operation requires multiple network transactions, greatly slowing the rate of computation.
Research into developing inefficient computational algorithms is relatively mature and focuses on the application of near-infinite loops (NIL's) to traditional computational techniques[4]. Applying these techniques to our users' codes has in some cases resulted in many orders-of-magnitude decreases in speed, down to the milliflop range.
The development of inefficient networking algorithms is also relatively well-understood, as exemplified by the well known software engineering practice of the Virtually Endless Network Transaction (VENT), in which "ping" requests and their corresponding acknowledgements (ACK)'s are implemented using NIL's[5].
However, a truly slow computer network is a vision further removed from reality than one may think. Acoustic networks reached a pinnacle with the development of the acoustically-coupled modem, which was able to achieve rates in the 100 baud (bits per second) range. However, telecommunications deregulation and the advent of variously-shaped handsets doomed this promising technology.
Optical networks such as the MPPA's face a daunting challenge in the speed of light, which in combination with its relatively short wavelengths allows for extremely high data transfer rates. Attempts by MPPA researchers to mitigate this issue by placing smoke machines in the room with the MPPA or putting sunglasses over the infrared transmitters resulted in lowering the reliability of the network, but not altering transfer rates.
Linking the Palm processors together with fiber-optic cables that run up and down the street a couple of times, forcing the light to travel farther, also had little effect. It was later computed by our financial office that we would have had to spend $30B on fiber-optic cable to achieve a significant slowdown using this techniqueIt was thus with great excitement that NCSA engineers read in Time magazine about recent research into lowering the speed of light[6]. No other single technique would do more to reduce the data-carrying capacity of optical networks. With slow enough light, transfer rates could plunge into the microbaud or nanobaud rates and result in most computational tasks being reduced to Near-Endless Wait States (NEWS's)[7].
No less than two techniques for slowing light have been discovered recently, both of which involve exciting-sounding lasers. We are still consulting with our financial office about paying for the necessary lasers.
This research was supported by the National Science Foundation's Partnership for Advanced Computational Inactivity and by the DOD's Security Through Obscurity program. Additional funds were provided by our financial department, which got them somewhere.
[1] F. Steckling. We Discovered Alien Bases on the Moon. G.A.F. International, 1997 .
[2] "Circuit City Home Page". http://www.circuitcity.com/
[3] Steele, Joy, and Wirth. The Fermata Programming Language, 3,482,754th Edition. MIT Press, 1985.
[4] J. F. Adams. "Near-Infinite Loop Spaces on a Budget". Taunton's Fine Mathematics, Summer 1982.
[5] J. Gray. "Sample and Hold: New Algorithms for Relational Database Mismanagement Systems". Proc. of ACM Subcomputing '94.
[6] "Bringing Light to a Stop". Time, 29 January 2001.
[7] S. Beckett. Waiting for Godot. Grove, 1997.