Hello, A yearly allocation for the LCRC cluster has been requested with the following updated information: Submitter/PI: Boyana Norris Project Name: pflotran Division: MCS Project title: PFLOTRAN Performance Analysis and Tuning Associated funding: SciDAC Institute for Sustained Performance, Energy, and Resilience (SUPER) Algorithms and Software for Communication Avoidance and Communication Hiding at the Extreme Scale (CACHE Math/CS Institute) Other Systems: Jazz, Surveyor, Intrepid Science: performance analysis, performance optimizations, and automated empirical tuning Project description: PFLOTRAN is a massively parallel code developed at LANL for the accurate prediction of multiscale subsurface processes. PFLOTRAN has many applications and guides safer industrial techniques; for example, it can help geoscientists determine how carbon dioxide flows through aquifers and deep geologic formations during carbon-capture and storage. In the first years of this project, we focused mainly on analyzing and improving the performance of the sequential portions of key computational kernels (matrix-vector product, triangular linear solution, and some of the chemistry computations), which required many single-core runs of these kernels. During FY12, we will focus on multicore optimizations and extensions of PETSc (used by PFLOTRAN) through our collaboration with RNET Technologies and Ohio State University. As part of the project, the team is investigating data structures, algorithms, compiler optimizations, and novel programming paradigms to enable applications based on the PETSc library to fully utilize the computing power of emerging Petascale compute architectures. While our initial focus is on PFLOTRAN, these modifications facilitate aggressive vectorization and efficient memory accesses and are expected to speed up many PETSc based finite difference and finite volume simulations. The project involves active development and testing of MPI, OpenMP, and manually created or auto-generated architecture intrinsics code. A parallel computing cluster with state-of-the-art architecture is necessary to demonstrate scalability and asymptotic performance of these PETSc modifications. Access to a cluster with multiple nodes, quad core processors, and high-performance interconnects is highly desirable in light of the above efforts. Another new focus for FY12 is dynamic adaptation of the linear solution methods by employing newly developed feature analysis and classification techniques. Based on analysis of limited training data, we identify significant matrix properties, which are then used to select the best solver and preconditioner combination in production runs of PFLOTRAN (and eventually other PETSc based applications). We also plan to continue our work on automatic code generation and tuning for the key linear algebra kernels in PFLOTRAN, which requires empirical testing of many different optimized versions. We will significantly extend the types of optimizations considered, which would correspondingly greatly increase the empirical search space, necessitating advances in the numerical optimization algorithms used to guide the search. We will use PFLOTRAN input decks that can run on different ranges of cores, from one node to the whole machine. We'll target run times of less than an hour in most cases (by changing the number of time evolution steps). Project URL: http://trac.mcs.anl.gov/projects/performance/wiki/OrioPFLOTRAN Current FY Hours Used: undetermined amount New FY Requested allocation: 499900 Q1: 124900 Q2: 125000 Q3: 125000 Q4: 125000 Justification: Thank You, The LCRC Accounts System