Frequently, applications that require several hours to a few weeks to execute on typical desktop computers are ported to supercomputers or large compute servers. Not only is this option costly, but often the entire program must be rewritten to effectively utilize multiple processors and other resources available on these systems. Reconfigurable computing (RC) systems can offer an advantage over this option since RC systems have been demonstrated that deliver performance approaching that of a supercomputer at a fraction of the cost. However, not only could it take several months to complete an RC system design, but RC system developers must be trained in both hardware design and software development.
The research project will develop a Remote Adaptive-Computing REsource on the Internet (RARE) that will assist scientists and engineers in solving problems that require excessive execution time when solved on a typical desktop computer. RARE comprises a comprehensive set of routines that have been optimized for high performance on a reconfigurable computer. A novel client/server program interface, developed as a part of this research effort, will manage the loading and execution of user process requests. The user interface to the resource consists of a library of functions and procedures that perform typical scientific/engineering computations. The goal of the project is to speedup local and remote versions of scientific data processing applications by orders of magnitude at reduced costs. Users of the system need only call a local or remote procedure that executes the application on the reconfigurable computing resource. A novel feature is that data is sent to RARE from a remote site, processed on hardware that has been optimized to perform a particular task significantly faster than a typical desktop computer, and transmitted back to the user via the Internet.
Recent experiments verify that remote hardware versions of several algorithms (implemented on a reconfigurable computer) can complete execution significantly faster than logical software versions of the same algorithm (implemented on a typical desktop computer). This statement is true in spite of the communication overhead that is present due to data transmission via the Internet.
Applications in the area of Computational Fluid Dynamics (CFD) are prime candidates for this resource since many of these computer simulation models require enormous execution times. Typically, CFD simulations require the solution of a large number of simultaneous equations or extensive computation of derivatives and/or integrals. One such CFD model, developed at NC State, requires two weeks to run on a desktop computer and this model is only a subset of the complete system simulation. Software profiling reports obtained during execution of these models demonstrate that more than 70% of the total execution time is spent in a small number of atomic procedures. Hence, these applications are the primary focus of this research project.
