Hello, A yearly allocation for the LCRC cluster has been requested with the following updated information: Submitter/PI: Michael Minkoff Project Name: IMLS_Methods Division: CSE Project title: Data Representation via Interpolative Methods for Massively Parallel Applications Associated funding: Other Systems: Clusters in the CSE's TCG group. Science: Based upon prior test software we have developed for calculating Potential Energy Surfaces (PES) using Interpolative Moving Least-Squares (IMLS) Methods, we are modifying our methods for data compression applications. We use interpolative methods like IMLS to represent elements in a data set to achieve a selected accuracy with fewer data elements than contained in the original data set. We have achieved data compressions of over a factor of a 1500 on data sets of twenty megabyte sizes. We need to scale the method up to terabyte data sets distributed over thousands of processors. Project description: We have developed a compression code based on multi-grid sampling of the data set. During this past year we have extended the software by use of MPI with the goal of optimizing performance. We identified three key kernels of the calculation that were suitable for parallelization and dominated the calculation time: i. Assembly of the matrix and vector of the least-squares calculation ii. Use of local grid optimization methods for identifying the best location for adding new basis points at each iteration iii. Optimization of the interpolation of the fit at the current grid to the next finer grid. We have successfully parallelized these three kernels and developed initial benchmarking results for a very limited number of test cases on up to 64 cores. Results show good speedup and high compression for problems of sufficient size. Using the requested resources, in this coming year we intend to do three things: 1. Further refine the parallel code via tests with a much larger test suite of compression applications at the ~100 core level. 2. Extend the size of each application to allow the determination of performance characteristics beyond the 1000 core level. 3. Develop an efficient the parallelization strategy for the least squares operation not included in the three kernels above. We do not intend to develop from scratch a parallelized least squares code but use existing parallelized codes. However, successful compression inherently means the least square operation processes much less data than the key kernels above. How to incorporate the least squares step with this mismatch (as a separate job assigned to fewer cores) will be our concern. We plan to bring our code to a plateau of development where its potential of extension of exascale application can be more reliably assessed. Project URL: Current FY Hours Used: undetermined amount New FY Requested allocation: 10000 Justification: Thank You, The LCRC Accounts System