An Object-Oriented Simulator for the Apiary

This paper describes a simulator for the proposed Apiary, an object-oriented, message passing parallel machine for artificial intelligence applications, using the QCKV model of computation. The simulator implements an interpreter for the lowest level “virtual machine language” of the Apiary, specifying computations in terms of creating objects and sending messages rather than loading and storing registers. The simulator is itself programmed in the object-oriented style advocated by the actor philosophy, allowing experimentation with alternative implementation mechanisms without disturbing the behavior of the simulation. Technical details in the paper assume some familiarity with object-oriented programming and the actor formalism. Paper category: Support Software and Hardware 1. Should a parallel machine for AI be like a parallel machine for nhvsics? What does it mean to build a machine optimized for artificial intelligence? Let’s look at the process of building specialized machines in other domains. Mathematics and physics, like AI, are areas which have important problems where solutions are limited by constraints on computing power. In these areas, an accepted methodoiogy for optimizing machines involves identifying the inner loop of some intcrcsting problem, a small piece of code that takes a large percentage of computing resources. Then. this inner loop is impleln~nted at as low a level as is feasible. preferably in microcode or directly in hardware. If what’s t&ins the time in your problem is doing FFT’s, build an F‘F7‘ machine. Can Al use this approach? Probably not. We conjcc,ture that AI doesn’t have a simple “inner loop”, that an Al machine will have to be a f%t “general purpose” problem solver. just ::s p:of.lc :~i’c. ‘I lrc tlil‘fcrcnce is that in physics problems the pa~~crns 01‘ LoIllpul:llion tend to be st,ltic and yredrcrtrhie. where:i< in Al the patterns of computation arc likely to be dynamic and thcrcfbre un~~l-t~~rlicl~zDI. An AI program attempting to solve a problem may have no idea which one of a number of heuristics will be useful before it starts to work on the problem. It may even have to learn or invent new solution methods as it goes along. Some specialized algorithms will undoubtedly be useful. such as pattern matching, set intersection and searching, but probably no one algorithm will be so dominant as to warrant tuning an Ai machine to just that algorithm. So what can you do to optimize a machine for unpredictable computations? First, you optimize the machine to take advantage of large amounts of parallelism. It will soon be more important to take advantage of the potential parallelism in a computation than to minimize the number of machine cycles used by a computation. It is important to optimize for flexibility, avoiding any sort of centralized control which might become a bottleneck. A consequence is that all resources in the machine should be allocated dynamically, including both memory and processor resources. Work should be distributed among parts of the machine as evenly as possible, to take maximum advantage of parallelism. Rather than dedicating special purpose hardware to particular algorithms, it is preferable to have many general purpose processors able to run parts of algorithms as the need arises. Computations should be able to move from processor to processor, even while they are running. Stored objects should be able to move from the memory of one processor to the memory of another processor without affecting programs that use the objects. The programmer should be able to program the machine pretending that an “infinite” number of processors are available, just as garbage collection and virtual memory let the programmer pretend an “infinite” number of memory cells arc available. The system should time-share available physical processors, just as virtual memory systems time-share the use of physical memory. Simple allocation strategies with good average behavior [like the least-rcccntly-used paging algorithm] should bc used to manage resource allocation. These are the design principles that serve as our criteria for a parallel machine for Al. The actor InCJdd of compu&tion, dc::cr-ibcd in [I]. 1.11. [,lJ. 151 1 7rovrdcs ;I 17~~s for dtGgning a machine which will meet thcsc criteria. 241 From: AAAI-83 Proceedings. Copyright ©1983, AAAI (www.aaai.org). All rights reserved.