No abstract

P arallel computing has emerged as a costeffective means of dealing with computationally intensive financial and scientific problems. To effectively utilize this technology, developers need software that reduces the complexity of the process as well as tools to support integration of parallel and desktop machines.The Clustertech parallel environment (CPE) is a C++ library that facilitates development of largescale parallel applications, particularly financial engineering applications. Written with performance and portability in mind, CPE currently runs on the Unix, Linux, and Windows operating systems.CPE provides domain-specific object-oriented libraries for solving partial/stochastic differential equations using the finite-difference method and Monte Carlo simulation. These libraries factor out the common operations required for FD and MC computations so that in most cases the user need only provide the code required for the specific application.CPE hides parallel synchronization and communications, allowing the user to emulate conventional serial programming; it also offers users better control of parallelization by overriding default methods. The domain-specific libraries are built on top of a set of high-performance parallel programming classes that ensure efficient communications and control.Although researchers have developed specialized parallel libraries for solving partial differential equations, 1 we are unaware of any other object-oriented parallel libraries for financial engineering applications that offer CPE's features. Sophisticated users may elect to execute programs directly on the parallel platform, but most commercial applications require integration of the parallel routines with existing software on the user's desktop machine-for example, an Excel spreadsheet, a Web-based interface, or a custom program. CPE provides mechanisms to seamlessly call and control parallel computations remotely from a desktop machine and transfer data within and between parallel machines and the desktop.CPE introduces several abstractions to simplify parallel application development. At the lowest level, a Tx class and related drivers unify communications, easing the task of transporting complex objects over different protocols. The MC implementation employs policies 2 to provide flexible control of the execution of these simulations. A distributed grid together with expression templates facilitate the implementation of partial differential equation solvers, allowing efficient manipulation of entire parallel grids using simple operators. Finally, remote execution enables the creation and manipulation of parallel objects from a desktop machine. CPE ARCHITECTUREAs much as possible, CPE decouples parallelization from the problem description. The libraries support both serial and parallel execution from the same source code so that in most cases users can The Clustertech parallel environment is an object-oriented C++ library that uses abstractions to simplify parallel programming for financial engineering applications....