Iaakov Exman

Solving Sequential Games with Boltzmann-Learned Tactics

A new approach is proposed to study strictly sequential problems viz. to substitute time sequences by ensembles of states, to be attained and viewed in parallel. 2-player board games are candidate problems for parallelization. Typical sequences of interest are game tactics to be learned. The latter are observable in the final board — the parallel view — of that sequence. We introduce parallel games, played in steps of m simultaneous moves. Non-deterministic relaxation rules are formulated, with Boltzmann factors, to resolve potential conflicts. An alternative way is simulated annealing to a sequence final board, as boards resemble spin systems. The approach enables a variety of novel game experiments. Sample results for specific tactics are discussed for parallel tic-tac-toe and othello.

The parallel break construct, or how to kill an activity tree

Most parallel languages provide means to express parallelism, e.g. a parallel-do construct, but no means to terminate the parallel activities spawned by such constructs. We propose three high-level primitives for this purpose, which are defined by analogies with primitives that break out of sequential iterative constructs. The primitives are pcontinue, which terminates the calling activity, pbreak, which terminates all the activities in the construct that spawned the calling activity, and return, which terminates all the activities created in the current function call. These constructs are especially useful in search problems, where an activity that finds a solution can terminate other activities that are investigating inferior approaches. Given that parallel constructs can be nested, activities form a tree rooted at the original activity that started the program. The main challenge in implementing pbreak and return is identifying the subtree of activities that should be killed. Three algorithms were designed and implemented, and experiments show that using these constructs can provide significant performance benefits.

Parallel Crystal Growth for Chip Placement

Parallel Crystal Growth, a new parallel extension of Simulated Annealing (SA), is presented. Sub-problems — the crystals — are solved in parallel, and iteratively merged into bigger crystals. The essential novelty is that processors spend time in a specific sub-problem, only as long as it is worthwhile, by locally computed merging criteria. Intra crystal search, is not repeated. Only inter-crystal exploration is done after merging. Whole problem termination is similarly determined. This leads to controlled speedup super-linearity, as demonstrated by experimental results with the PVM implementation.

Parallel roadmaps: discrete to continuous

Parallel Roadmaps are apparently naive visual constructs, which remain intelligible and provide invaluable debugging clues as program complexity increases. With a few activities, roadmaps are discrete and detailed. With large numbers of them, continuous regions are formed by coalescing of adjacent activities. The main thesis of this paper is that a smooth transition from discrete to continuous maps, without changing the basic graphical entities, is an effective means to deal with parallelism at largely different scales.<<ETX>>

On-the-fly message patching in free-space optical interconnects

Free-space optical interconnects are exploited to solve general purpose computing tasks which remap structured sets of information. These tasks -- viz. load balance, monitoring, fault tolerance, shadowing -- can all be based upon elementary message patching operations. It is shown that usual assumptions in optical computing -- the redundancy of numbers, as in dual rail representation -- suffice to perform recoding, thus patching, in elegant ways. Smart recoding schemes, may be even less redundant than dual-rail. For monitoring of large-scale computer systems, small numbers of smallitors are added/removed as needed in between processor clusters. Smallitors Snatch-&-Patch messages, by intercepting and partially deflecting optical information. Smallitors are small in size and in complexity. Once one is given enough bandwidth, the next critical limiting resource is latency. This is a prime reason behind the on-the-fly patching approach in free-space optics. Smallitors strive to minimize latency, by turning on additional bits in parallel, thereby recoding the transmitted ones. Recoding schemes and their optical implementation, are discussed within concrete computing tasks. Observations from actual prototypes are also considered.

Envelopes in Adaptive Local Queues for MIMD Load Balancing

Envelopes, a run-time mechanism which automatically supports adaptive local queues for MIMD load balancing, are proposed and demonstrated. Envelopes promote generality and language simplicity, while sustaining efficiency.

Optical Interconnects for Parallel Systems: Demonstration on an Optical Link with Multiple-Quantum-Well Opto-Electronic Arrays

Implementations of massively parallel computer systems suffer from the well known communications bottleneck existing between different boards as well as racks. Optics may be an alternative by providing parallel interconnects with high bandwidth and density. We investigate an optical link using two bidimensional multiple-quantum-wells (M.Q.W) opto-electronic arrays, used reversibly as spatial light modulators and receptors. In a first step, we implement a 4×4 channels demonstrator at 30 MHz rate.