KBLAS (KAUST BLAS) is a small library that provides highly optimized BLAS routines on systems accelerated with GPUs. KBLAS is entirely written in CUDA C, and targets NVIDIA GPUs with compute capability 2.0 (Fermi) or higher. The current focus is on level-2 BLAS routines, namely the general matrix vector multiplication (GEMV) kernel, and the symmetric/hermitian matrix vector multiplication (SYMV/HEMV) kernel. KBLAS provides these two kernels in all four precisions (s, d, c, and z), with support to multi-GPU systems. Through advanced optimization techniques that target latency hiding and pushing memory bandwidth to the limit, KBLAS outperforms state-of-the-art kernels by 20-90% improvement. Competitors include CUBLAS-5.5, MAGMABLAS-1.4.0, and CULAR17. The SYMV/HEMV kernel from KBLAS has been adopted by NVIDIA, and should appear in CUBLAS-6.0. KBLAS has been used in large scale simulations of multi-object adaptive optics.

European Extreme Large Telescope (E-ELT) is a high priority project in ground based astronomy that aims at constructing the largest telescope ever built. MOSAIC is an instrument proposed for E-ELT using Multi- Object Adaptive Optics (MOAO) technique for astronomical telescopes, which compensates for effects of atmospheric turbulence on image quality, and operates on patches across a large FoV.

Many real world networks have inherent community structures, including social networks, transportation networks, biological networks, etc. For large scale networks with millions or billions of nodes in real-world applications, accelerating current community detection algorithms is in demand, and we present two approaches to tackle this issue -A K-core based framework that can accelerate existing community detection algorithms significantly; -A parallel inference algorithm via stochastic block models that can distribute the workload.

Hierarchical matrix approximations are a promising tool for approximating low-rank matrices given the compactness of their representation and the economy of the operations between them. Integral and differential operators have been the major applications of this technology, but they can be applied into other areas where low-rank properties exist. Such is the case of the Block Cyclic Reduction algorithm, which is used as a direct solver for the constant-coefficient Poisson quation. We explore the variable-coefficient case, also using Block Cyclic reduction, with the addition of Hierarchical Matrices to represent matrix blocks, hence improving the otherwise O(N2) algorithm, into an efficient O(N) algorithm.

Among optimal hierarchical algorithms for the computational solution of elliptic problems, the Fast Multipole Method (FMM) stands out for its adaptability to emerging architectures, having high arithmetic intensity, tunable accuracy, and relaxed global synchronization requirements. We demonstrate that, beyond its traditional use as a solver in problems for which explicit free-space kernel representations are available, the FMM has applicability as a preconditioner in finite domain elliptic boundary value problems, by equipping it with boundary integral capability for finite boundaries and by wrapping it in a Krylov method for extensibility to more general operators. Compared with multilevel methods, it is capable of comparable algebraic convergence rates down to the truncation error of the discretized PDE, and it has superior multicore and distributed memory scalability properties on commodity architecture supercomputers.

Graphics Processing Units (GPUs) are gradually becoming mainstream in supercomputing as their capabilities to significantly accelerate a large spectrum of scientific applications have been clearly identified and proven. Moreover, with the introduction of high level programming models such as OpenACC [1] and OpenMP 4.0 [2], these devices are becoming more accessible and practical to use by a larger scientific community. However, performance optimization of OpenACC accelerated applications usually requires an in-depth knowledge of the hardware and software specifications. We suggest a prediction-based performance tuning mechanism [3] to quickly tune OpenACC parameters for a given application to dynamically adapt to the execution environment on a given system. This approach is applied to a finite difference kernel to tune the OpenACC gang and vector clauses for mapping the compute kernels into the underlying accelerator architecture. Our experiments show a significant performance improvement against the default compiler parameters and a faster tuning by an order of magnitude compared to the brute force search tuning.

This research aims to understand the performance of PETSc-FUN3D, a fully nonlinear implicit unstructured grid incompressible or compressible Euler code with origins at NASA and the U.S. DOE, on many-integrated core architecture and how a hybridprogramming paradigm (MPI+OpenMP) can exploit Intel Xeon Phi hardware with upwards of 60 cores per node and 4 threads per core. For the current contribution, we focus on strong scaling with many-integrated core hardware. In most implicit PDE-based codes, while the linear algebraic kernel is limited by the bottleneck of memory bandwidth, the flux kernel arising in control volume discretization of the conservation law residuals and the preconditioner for the Jacobian exploits the Phi hardware well.