Parallel Infrastructure for Recognition of Non-Local Patterns from Particle Detectors

点击翻译按钮获取中文摘要

This project studies pattern recognition algorithms for High Energy Physics(HEP). In particular, it looks at pattern recognition for data from HEPexperiments which requires global information and lends itself well toparallelism. While a trivial form of parallelism has always been presentin HEP software due to the independent data (results of separate highenergy collisions), there has been little done for pattern recognitionwithin single events. This is a particularly pressing problem for the nextgeneration of experiments where data is taken in at rates of up to 20Mbytes each 25 nanoseconds and rapid pattern recognition is essential todecide which events should be kept for later study. Since a single eventwill leave traces in many different detectors, pattern recognition here isa non-local, non-sequential pattern matching problem.Other physicists have tried to find a quick heuristic match during a singlepass through the data. In contrast, this project will do a more carefulanalysis of the data to better resolve multiple events that are close intime and space by using signal processing techniques (e.g. Kalmanfiltering) to pre-process the data. The project will use a betteralgorithmic formalism to take advantage of computer science algorithms anddata structures, and will implement parallel algorithms to recognizenon-local patterns. Special challenges will be presented by theanticipated datasets, which have petabyte sizes. New algorithms for theseapplications will have to take careful account of the different accesstimes for RAM and disk access. This project is interdisciplinary innature, with potential impact on both computer science and high energyphysics.