Even a modest lattice QCD project demands quite large amount of computing resources. In this regard, it has always been an attractive idea to build a cheap high performance computing platform out of commodity PC’s and commodity networking devices. However, the availability of cheap hardware components solved only a part of problem in building a parallel computer in the past. There were large hi...

On cluster-based systems, data is typically stored on a centralized resource, and each node has a local disk used for the operating system and swap space. Although I/O middlewares (e.g., MPI-IO) and high performance I/O subsystems (e.g., RAID) can generate parallel I/O streams, disk contention and network latency still dominate I/O performance. To address this performance barrier, I/O access ne...

This paper describes the development of the PALS system, an implementation of Prolog capable of efficiently exploiting or-parallelism on distributed-memory platforms— specifically Beowulf clusters. PALS makes use of a novel technique, called incremental stack-splitting. The technique proposed builds on the stack-splitting approach, previously described by the authors and experimentally validate...

One of the rst questions that a user of a new Beowulf-type system asks is, \How fast does my system run?" Performance is more than just a curiosity for cluster systems. It is arguably the central motivation for building a clustered system in the rst place|a single node is not su cient for the task at hand. Thus the measurement of performance, and comparisons of performance between various avail...