The Cosmic Microwave Background (CMB) is an exquisitely sensitive probe of the fundamental parameters of cosmology. Extracting this information is computationally intensive, requiring massively parallel computing and sophisticated numerical algorithms. In this work we present MAD bench, a lightweight version of the MADCAP CMB power spectrum estimation code that retains the operational complexity and integrated system requirements. In addition, to quantify communication behavior across a variety of architectural platforms, we introduce the Integrated Performance Monitoring (IPM) package: a portable, lightweight, and scalable tool for effectively extracting MPI message-passing over heads. A performance characterization study is conducted on some of the world's most powerful supercomputers, including the superscalar Seaborg (IBMPower3+) and CC-NUMA Columbia (SGI Altix), as well as the vector-based Earth Simulator (NEC SX-6 enhanced) and Phoenix (Cray X1) systems. In-depth analysis shows that in order to bridge the gap between theoretical and sustained system performance, it is critical to gain a clear understanding of how the distinct parts of large-scale parallel applications interact with the individual subcomponents of HEC platforms.

The Conjugate Gradient (CG) algorithm is perhaps the best-known iterative technique to solve sparse linear systems that are symmetric and positive definite. In previous work, we investigated the effects of various ordering and partitioning strategies on the performance of CG using different programming paradigms and architectures. This paper makes several extensions to our prior research. First, we present a hybrid(MPI+OpenMP) implementation of the CG algorithm on the IBM SP and show that the hybrid paradigm increases programming complexity with little performance gains compared to a pure MPI implementation. For ill-conditioned linear systems, it is often necessary to use a preconditioning technique. We present MPI results for ILU(0) preconditioned CG (PCG) using the BlockSolve95 library, and show that the initial ordering of the input matrix dramatically affect PCG's performance. Finally, a multithreaded version of the PCG is developed on the Cray (Tera) MTA. Unlike the message-passing version, this implementation did not require the complexities of special orderings or graph dependency analysis. However, only limited scalability was achieved due to the lack of available thread level parallelism.

Computational grids have the potential for solving large-scale scientific problems using heterogeneous and geographically distributed resources. However, a number of major technical hurdles must be overcome before this potential can be realized. One problem that is critical to effective utilization of computational grids is the efficient scheduling of jobs. This work addresses this problem by describing and evaluating a grid scheduling architecture and three job migration algorithms. The architecture is scalable and does not assume control of local site resources. The job migration policies use the availability and performance of computer systems, the network bandwidth available between systems, and the volume of input and output data associated with each job. An extensive performance comparison is presented using real workloads from leading computational centers. The results, based on several keymetrics, demonstrate that the performance of our distributed migration algorithms is significantly greater than that of a local scheduling framework and comparable to a non-scalable global scheduling approach.

Adaptive mesh refinement (AMR) is a powerful technique that reduces the resources necessary to solve otherwise in-tractable problems in computational science. The AMR strategy solves the problem on a relatively coarse grid, and dynamically refines it in regions requiring higher resolution. However, AMR codes tend to be far more complicated than their uniform grid counterparts due to the software infrastructure necessary to dynamically manage the hierarchical grid framework. Despite this complexity, it is generally believed that future multi-scale applications will increasingly rely on adaptive methods to study problems at unprecedented scale and resolution. Recently, a new generation of parallel-vector architectures have become available that promise to achieve extremely high sustained performance for a wide range of applications, and are the foundation of many leadership-class computing systems worldwide. It is therefore imperative to understand the tradeoffs between conventional scalar and parallel-vector platforms for solving AMR-based calculations. In this paper, we examine the HyperCLaw AMR framework to compare and contrast performance on the Cray X1E, IBM Power3 and Power5, and SGI Altix. To the best of our knowledge, this is the first work that investigates and characterizes the performance of an AMR calculation on modern parallel-vector systems.