Search

Grid Computing proposes unlimited access to different computational resources in a transparent way. High performance execution in grid environments is virtually impossible without timely access to accurate and up-to-date information related to distributed resources and services. Due to inherent difficulty of testing the different information propagation policies in real grid infrastructures, several simulation frameworks arose to help in this issue. In this work, we present Grid Matrix, an extension to one of the most used grid simulation tools (SimGrid2) to focus on the propagation of monitoring and resource information allowing the creation of virtual grid infrastructures. This extension enables GUI editing of network topology and provides the feature of scripting to define simulation details based on the newly developed C++ and Python bindings of SimGrid2 API. As a case study, Grid Matrix was used to test four different policies: hierarchical, super-peer, best-neighbor and random. The simulated scenario consisted of 96 master nodes based on the real Teragrid infrastructure as was publicly available at the time of writing this paper. We introduce three metrics that capture and summarize the information propagation behavior: LIR, GIR and GIV. LIR captures the local behavior quantifying the amount of up-to-date information in each node. GIR evaluates the global information state in the whole network averaging the LIR values, while GIV measures the variability of LIR. In the presented case, the best results in terms of the proposed metrics were attained by the hierarchical policy, followed by super-peer which outperformed random and best-neighbor. The modern and modular design of the scripting features included in Grid Matrix, in close conjunction with the user friendly GUI happened to be a very powerful tool for the evaluation of new propagation policies of resource information.

Models are increasingly being used to quantify the effects of best management practices (BMPs) on water quality. While these models offer the ability to study multiple BMP scenarios, and to analyze impacts of various management decisions on watershed response, associated analyses can be very computationally intensive due to a large number of runs needed to fully capture the various uncertainties in the model outputs. There is, thus, the need to develop suitable and efficient techniques to handle such comprehensive model evaluations. We demonstrate a novel approach to accomplish a large number of model runs with Condor, a distributed high-throughput computing framework for model runs with the Soil and Water Assessment Tool (SWAT) model. This application required more than 43,000 runs of the SWAT model to evaluate the impacts of 172 different watershed management decisions combined with weather uncertainty on water quality. The SWAT model was run in the Condor environment implemented on the TeraGrid. This framework significantly reduced the model run time from 2.5 years to 18 days and enabled us to perform comprehensive BMP analyses that may not have been possible with traditional model runs on a few desktop computers. The Condor system can be used effectively to make Monte Carlo analyses of complex watershed models requiring a large number of computational cycles.

Grid-computing environments are becoming increasingly popular for scientific computing due to the significant increase in capacity that they represent when compared to a single computational resource (e.g., a single cluster), their ubiquitous availability and advances in grid middleware components. Several such specialized grid environments are now available for users in the commercial and research sectors. One such effort for research is the TeraGrid, consisting of a collection of geographically distributed heterogeneous supercomputer resources including data storage resources. Parallel implementations for these environments are inherently multilevel and obtaining efficient mapping of work to processors can be eXtremely challenging. This paper eXtends an eXisting MPI application to the grid via the use of grid-enabled MPI libraries. The application uses a simulation-optimization framework involving coarse-grained parallelism in the optimizer and fine-grained parallelism in the finite-element-based simulator. Using parallelism at both these levels is essential for problems involving computationally intensive simulation steps. A hierarchical grid architecture consisting of a collection of supercomputers is ideally suited for these types of problems as a good application-to-architecture mapping can be obtained with a proper implementation. This paper presents the performance results of our implementation on the TeraGrid network consisting of three geographically distributed supercomputers.