Covariance matrices are ubiquitous in computational science and engineering. In particular, large covariance matrices arise from multivariate spatial data sets, for instance, in climate/weather modeling applications to improve prediction using statistical methods and spatial data. One of the most time-consuming computational steps consists in calculating the Cholesky factorization of the symmetric, positive-definite covariance matrix problem. The structure of such covariance matrices is also often data-sparse, in other words, effectively of low rank, though formally dense. While not typically globally of low rank, covariance matrices in which correlation decays with distance are nearly always hierarchically of low rank. While symmetry and positive definiteness should be, and nearly always are, exploited for performance purposes, exploiting low rank character in this context is very recent, and will be a key to solving these challenging problems at large-scale dimensions. The authors design a new and flexible tile row rank Cholesky factorization and propose a high performance implementation using OpenMP taskbased programming model on various leading-edge manycore architectures. Performance comparisons and memory footprint saving on up to 200K × 200K covariance matrix size show a gain of more than an order of magnitude for both metrics, against state-of-the-art open-source and vendor optimized numerical libraries, while preserving the numerical accuracy fidelity of the original model. This research represents an important milestone in enabling large-scale simulations for covariancebased scientific applications.

Abstract. For outer-product-parallel sparse matrix-matrix multiplication (SpGEMM) of the form C = A×B, we propose three hypergraph models that achieve simultaneous partitioning of input and output matrices without any replication of input data. All three hypergraph models perform conformable one-dimensional (1D) columnwise and 1D rowwise partitioning of the input matrices A and B, respectively. The first hypergraph model performs two-dimensional (2D) nonzero-based partitioning of the output matrix, whereas the second and third models perform 1D rowwise and 1D columnwise partitioning of the output matrix, respectively. This partitioning scheme induces a two-phase parallel SpGEMM algorithm, where communication-free local SpGEMM computations constitute the first phase and the multiple single-node-accumulation operations on the local SpGEMM results constitute the second phase. In these models, the two partitioning constraints defined on weights of vertices encode balancing computational loads of processors during the two separate phases of the parallel SpGEMM algorithm. The partitioning objective of minimizing the cutsize defined over the cut nets encodes minimizing the total volume of communication that will occur during the second phase of the parallel SpGEMM algorithm. An MPI-based parallel SpGEMM library is developed to verify the validity of our models in practice. Parallel runs of the library for a wide range of realistic SpGEMM instances on two large-scale parallel systems JUQUEEN (an IBM BlueGene/Q system) and SuperMUC (an Intel-based cluster) show that the proposed hypergraph models attain high speedup values. 1. Introduction. Sparse matrix-matrix multiplication (SpGEMM) is a kernel operation in a wide variety of scientific applications such as finite element simulations based on domain decomposition [3,22], molecular dynamics (MD) [15,16,17,25,28,29,32,36], and linear programming (LP) [7,8,26], all of which utilize parallel processing technology to reduce execution times. Among these applications, below we exemplify three methods/codes from which we select realistic SpGEMM instances.In finite element application fields, finite element tearing and interconnecting (FETI) [3,22] type domain decomposition methods are used for numerical solution of engineering problems. In this application, the SpGEMM computation GG T is performed, where G = R T B T , R is the block diagonal basis of the stiffness matrix, and B is the signed matrix with entries −1, 0, 1 describing the subdomain interconnectivity.In MD application fields, CP2K program [1] performs parallel atomistic and

No abstract

Abstract. Exploiting data sparsity in dense matrices is an algorithmic bridge between architectures that are increasingly memory-austere on a per-core basis and extreme-scale applications. The Hierarchical matrix Computations on Manycore Architectures (HiCMA) library tackles this challenging problem by achieving significant reductions in time to solution and memory footprint, while preserving a specified accuracy requirement of the application. HiCMA provides a high-performance implementation on distributed-memory systems of one of the most widely used matrix factorization in large-scale scientific applications, i.e., the Cholesky factorization. It employs the tile low-rank data format to compress the dense data-sparse off-diagonal tiles of the matrix. It then decomposes the matrix computations into interdependent tasks and relies on the dynamic runtime system StarPU for asynchronous out-of-order scheduling, while allowing high user-productivity. Performance comparisons and memory footprint on matrix dimensions up to eleven million show a performance gain and memory saving of more than an order of magnitude for both metrics on thousands of cores, against state-of-the-art open-source and vendor optimized numerical libraries. This represents an important milestone in enabling large-scale matrix computations toward solving big data problems in geospatial statistics for climate/weather forecasting applications.

Sparse matrix-vector multiplication (SpMxV) is a kernel operation widely used in iterative linear solvers. The same sparse matrix is multiplied by a dense vector repeatedly in these solvers. Matrices with irregular sparsity patterns make it difficult to utilize cache locality effectively in SpMxV computations. In this work, we investigate single-and multiple-SpMxV frameworks for exploiting cache locality in SpMxV computations. For the single-SpMxV framework, we propose two cache-size-aware top-down row/column-reordering methods based on 1D and 2D sparse matrix partitioning by utilizing the column-net and enhancing the row-column-net hypergraph models of sparse matrices. The multipleSpMxV framework depends on splitting a given matrix into a sum of multiple nonzero-disjoint matrices so that the SpMxV operation is performed as a sequence of multiple input-and output-dependent SpMxV operations. For an effective matrix splitting required in this framework, we propose a cache-size-aware top-down approach based on 2D sparse matrix partitioning by utilizing the row-column-net hypergraph model. The primary objective in all of the three methods is to maximize the exploitation of temporal locality. We evaluate the validity of our models and methods on a wide range of sparse matrices by performing actual runs through using OSKI. Experimental results show that proposed methods and models outperform state-of-the-art schemes.