The goal of the SIParCS program is to make a longterm, positive impact on the quality and diversity of the workforce needed to use and operate 21st century supercomputers. Graduate students and undergraduate students gain significant handson experience in highperformance computing and related fields that use HPC for scientific discovery and modeling. This program embeds students as summer interns in the Computational and Information Systems Laboratory, an organization within NCAR charged with provisioning supercomputing and data systems to the geosciences research community.
Pages


A comparison of NCAR application performance on IBM power architectures [presentation]

The performance of several of NCAR's climate models was measured on Bluefire and then compared with similar data from Blueice and Bluevista. Both Blueice and Bluevista are Power5 systems, whereas Bluefire is a Power6 system. Differences can be observed between the performance of these different architectures, and new technologies such as Altivec/VMX can be exploited on Bluefire. This work primarily relies on IBM's HPM toolkit to gather hardware counter data. From this counter data, differences in chip performance can be observed.


A conservative semiLagrangian discontinuous Galerkin method for transport equation on the cubedsphere [video]

The discontinuous Galerkin (DG) method is becoming increasingly popular in atmospheric and ocean modeling. However, a major drawback of method is its stringent CFL stability restriction associated with explicit timestepping, e.g. explicit RungeKutta method. In order to get around this issue we adopt a dimensionsplitting approach where a regular semiLagrangian (SL) scheme is combined with the DG method. The resulting SLDG scheme employs a sequence of 1D operations for solving transport equation on the cubedsphere. The SLDG scheme is inherently conservative and has the option to incorporate a local positivitypreserving filter for tracers. A novel feature of the SLDG algorithm is that it can be used for multitracer transport for global models employing spectralelement grids.


A coordinate based semiLagrangian algorithm [presentation]

In this talk it is shown how an advection algorithm (available in a dycore for example) can be used to create a semiLagragian algorithm of arbitrary order of accuracy in time and space using operator integrating factor splitting (OIFS). This algorithm can potentially transport hundreds of tracers with only kinterpolation(s) per tracer per timestep to yield kth order accuracy in time. The method is more efficient than advecting the the tracers individually after some small number of tracers has been exceeded. This threshold decreases with decreasing order in time and increasing polynomial order and is bounded below by the number of coordinates advected (6 for a sphere). Preliminary results in 1D and 2D are presented where the underlying discretization scheme is the discontinuous Galerkin method (DG) in strong and weak form.


A fluxform version of the conservative semiLagrangian multitracer transport scheme (CSLAM) [presentation]

We present a fluxform modification of a semiLagrangian advection scheme that computes the flux of an advected quantity for use in a finitevolume method. The fluxform scheme adds little overhead compared to the semiLagrangian method, while also allowing the use of flux limiters to ensure monotonicity of the solution. Using flux limiting is seen to allow better accuracy and convergence rates than monotonicityensuring schemes available to semiLagrangian schemes, while being much more efficient.


A hybrid estimator for density with extremes [presentation]

Extreme weather events that are often important in assessing the impact of climate on our society and the natural environment. To model a nonparametric long tailed distribution, the developed kernel density and logspline method may lack of accuracy in the tails, while the General Pareto Distribution (GPD) do not give the information to the nonextreme section of the data. This study is about a hybrid statistics model combining the logspline method with a parametric GPD tail, by which we can learn about the whole density of the variable and give a more accurate estimation on the probability of the extreme events.


A nonstaggered block Jacobi preconditioning strategy in HOMME [presentation]

Due to parallel implementation, the original block Jacobi preconditioner without any communications for staggered grid in HOMME doesn't work for the nonstaggered grid. This report will introduce why the original preconditioner fails for the nonstaggered case, the challenges in implementing it, and the basic idea of the algorithm of the nonstaggered preconditioner. Comparison with the original preconditioner and analysis will also be addressed at the end.


A twolevel nonoverlapping optimized Schwarz for spectral elements [presentation]

Semiimplicit timestepping yields a positive definite Helmholtz problem that needs to be inverted at each timestep. When combined with a semiLagrangian approach, the resulting Helmholtz problem can potentially become stiff and the number of iterations required to invert the problem grows without bounds. When discretized in space using spectral elements, it is possible to devise a nonoverlapping Schwarz algorithm that uses optimized transmission operators between subdomains. However, to take care of the stiffness, a coarse solver is required. We report recent results for the Poisson and positive definite Helmholtz operators discretized using 1D and 2D spectral elements.
Pages