Mutual algorithm-architecture analysis for real-time parallel systems in particle physics experiments.

摘要

Data acquisition from particle colliders requires real-time detection of tracks and energy clusters from collision events occurring at intervals of tens of {dollar}mu{dollar}s. Beginning with the specification of a benchmark track-finding algorithm, parallel implementations have been developed. A revision of the routing scheme for performing reductions such as a tree sum, called the reduced routing distance scheme, has been developed and analyzed. The scheme reduces inter-PE communication time for narrow communication channel systems. A new parallel algorithm, called the interleaved tree sum, for parallel reduction problems has been developed that increases efficiency of processor use. Detailed analysis of this algorithm with different routing schemes is presented. Comparable parallel algorithms are analyzed, also taking into account the architectural parameters that play an important role in this parallel algorithm analysis. Computation and communication times are analyzed to guide the design of a custom system based on a massively parallel processing component. Developing an optimal system requires mutual analysis of algorithm and architecture parameters. It is shown that matching a processor array size to the parallelism of the problem does not always produce the best system design. Based on promising benchmark simulation results, an application specific hardware prototype board, called Dasher, has been built using two Blitzen chips. The processing array is a mesh-connected SIMD system with 256 PEs. Its design is discussed, with details on the software environment.