Aximetric and the Cornell Theory Center (CTC) strongly backed by Microsoft High Performance Computing (HPC) lab proposes a portable parallel computing model for windows platform called Java based Bulk Synchronous Parallel System (JBSP) that is implemented using the .net clustering technology developed by Microsoft. The uniqueness of this approach is to provide portable model called JBSP.Net using the BSP cost model along with Java like visual environment provided by Microsoft. Current approaches for portable parallel computing will be studied and proposed in light of the Sensory fusion problem, and new approaches for load balancing using the proposed model will be presented. A graph-based load balancing system (GPSys) will be developed to assist in load balancing and performance mapping. The Bulk Synchronous Parallel (BSP) model (Valient, 90) is a very portable model that uses simple programming parameters for building the parallel programming model. These models can be built on uniprocessor machines and the same can be easily transferred to parallel/distributed processors using the BSP library for those platforms. The compile once run anywhere concept of Java, the BSP parallel model and the newly developed .Net architecture (Microsoft, 02) developed by microsoft to network servers, form the crux of our approach towards portability and scalability using low cost windows workstations. Aximetric team is proposing a unique graph based architecture to assist in building the portable parallel model. The BSP model similar to LogP model is portable, yet lacks tools to provide measures of latency, blocked times and parallel slackness with respect to the problem modeled. Portablity has to balanced with efficiency of processors and interprocessor communications. It is well known that in general optimal multiprocessor mapping problem is NP-hard. The processor allocation problem can be viewed as a weighted graph mapping, where the nodes of a source graph are mapped into those of a target graph. In the source graph the nodes represent computational tasks and the edges represent the inter-process communication. The target graph represents the parallel architecture. The objective function, which models the cost associated with the mapping, is to be minimized. We will use the agencys Sensory fusion applications for testing and benchmarking the proof of concept. Aximetric team will also use a unique tagging and segmenting approach for the data tree that can be used for storing sensory data. This approach brings design, development, testing, running and visualization for parallel programs in a common environment used in main stream computing. The application will be tested on CTC's 256 node Dell Windows 2K cluster. The objectives of the Phase I effort are to define the user requirements, identify and define the critical design attributes, and demonstrate the feasibility of the portable parallel model concept for sensory problems and alike using the .Net cluster. During Phase II, a full-scale prototype will be developed. This prototype will be subjected to a comprehensive field test at a predefined DoD center and revised as appropriate. A marketing plan will be developed so that the software can be commercialized and introduced to the marketplace during Phase III. As more techniques to load balance, characterize the cost model, meta-tag data and partition data are developed, the concept can be leveraged to any intelligent data mining needs throughout the intelligence and business community, Environmental Modeling for real-time simulation and visualization, in the government and industry regardless of subject matter content.
Keywords: Distributed Computing, Parallel Computing, Jbsp.Net, Bulk Synchronous Parallel Model (Bs, .Net Cluster, Logp, Intelligent Computing
