Student Posters Repository

The use of adjoint-based error estimation in conjunction with a highly parallel and scalable, anisotropic, block-based, adaptive mesh refinement (AMR) technique is considered for the more efficient prediction of three-dimensional compressible flows. In particular, a comparison is made between the computational performances of output-based error estimates, derived via two approaches, namely one based on mesh (or h) refinement and the other based on order (or p) refinement, for directing the mesh refinement in anisotropic AMR scheme. The AMR scheme allows enhancement of local mesh resolution, with preference given to directions as dictated by the flow solution. The proposed adjoint-based error estimation technique provides a posteriori estimates of the error for an engineering functional of interest in terms of estimates of the local solution error following from the solution residual. The estimated error in the solution residual is obtained either via direct refinement of the mesh in the preferred directions (here referred to as the h-derived error indicator) or by using a higher-order spatial operator with anisotropic feature detection based on the Hessian of an appropriate solution quantity (here referred to as the p-derived error indicator). Both approaches are considered here. Additionally, two formulations of the adjoint-based error indicator are examined here for directing the output-based AMR. The first is the so-called computable correction (CC), where the residual error is weighted by the corresponding adjoint solution for the functional of interest, and the second is the so-called error in the computable correction (ECC), which is comprised of a linear combination of the residual error weighted with the adjoint solution and the adjoint residual weighted with the primal solution. The resulting output error indicator is used to direct the mesh refinement, with regions of the solution domain contributing most significantly to the functional error being selected for local enrichment of the mesh. In this way, the computed accuracy of the functional is increased while potentially greatly reducing the associated computational cost of performing the simulation. For the cases of interest, both low- and high-order upwind finite-volume spatial discretization schemes are applied in conjunction with the block-based AMR scheme to the solution of the partial differential equations governing steady-state inviscid compressible flows. The potential benefits of the proposed anisotropic block-based AMR with adjoint-based error estimation are demonstrated for a range of compressible inviscid flow problems of varying complexity. Comparisons of solution accuracy and relative computational costs for results obtained using both h- and p-derived error estimates of the solution residual are examined and discussed.

The satellite Solar Dynamics Observatory observes the atmosphere of the Sun in the extreme ultra-violet (EUV) at a high resolution (4kx4k) and a high cadence (12s). The continuous observations allows us to track solar coronal holes (CH), i.e., regions in the solar corona with a reduced density, temperature, and an "open" magnetic field topology. Further, these regions are the source of high-speed solar wind streams. Whenever these high-speed plasma streams hit the Earth, they cause geomagnetic storms, which are responsible for polar lights, an increased drag-force on satellites, and geomagnetically induced currents (GIC). Here, we present the revised coronal hole module of the SPoCA suite, which allows unsupervised identification and tracking of coronal holes.

High-fidelity bio-simulation tools have been applied in computational medicine to decipher the underlying cellular-level mechanisms of pathophysiology, predicting treatment response and suggest the most pertinent personalized treatments. To gain accuracy by increasing the relevant biological resolution, speed is often compromised as the model produces substantial amount of data in each iteration. This work presents a high-performance scheme designed to organize subtasks of agent-based model (ABM) of vocal fold inflammation and repair to fully utilize a heterogeneous platform consisting of multi-core CPU host with attached GPUs. The subtasks are further optimized based on their execution and data access characteristics. Conjointly, with the employment of VirtualGL, the model visualizes the results in situ providing the user with computational steering capability and interactivity in real-time. This 3D biological model is capable of processing (compute and visualize remotely) 154 million vocal fold biomarker data points per second, which is 900x more powerful than popular existing ABMs framework (NetLogo).

Ensemble Kalman filter (EnKF) is an algorithm of numerical weather forecast (NWF) which combines numerical simulations and various types of observations interactively. The weather system is incredibly high-dimensional space. EnKF assumes the probability density function (PDF) of the current state of systems, e.g. temperature, moisture, and wind velocity field, to be Gaussian to reduce the calculation cost due to the high-dimensionality. This approximation works well for global forecast, however, it seems bad for local severe event, where instantaneous local nonlinear dynamics are important. This study evaluates the non-Gaussian effect as a bias of ensemble estimations. The non-Gaussian bias is revealed to be state-dependent, and is critically large only on a limited case. This study is done only on a low-dimensional system, and will be extended to realistic systems using HPC techniques.

The Canadian Hydrogen Intensity Mapping Experiment (CHIME) is one of a new class of cosmology experiments, mapping redshifted neutral hydrogen throughout the universe to learn more about one of the great scientific questions of our time: the accelerating expansion of the universe, i.e. dark energy. CHIME is a radio transit telescope consisting of four 20 m x 100 m cylinders with over 1000 antennas located in British Columbia, Canada. When completed later this summer, CHIME will generate about 135 TB of raw data per day, which must be carefully calibrated and compressed for storage. One tool CHIME can use for both calibration and compression is redundancy within the array, the property that many data products should contain the same information once calibrated. Harnessing this redundancy and accounting for imperfections in the telescope generate interesting analytical and computational challenges. 

MPI libraries today are complex software's that have modular components reflecting the constantly evolving software and hardware stack underneath. They offer the user a number of tunable parameters that can affect performance. In this environment, performance variations and scalability limitations can come from a number of sources. It is thus crucial to gain a holistic understanding of MPI library performance in order to fine-tune it for a given application and hardware setup. The MPI Tools Information Interface (MPI_T) introduced as part of the MPI 3.0 standard offers a standardized mechanism through which external tools can interact with an MPI library to perform runtime introspection and performance tuning. In this work, we describe a plugin infrastructure that utilizes the MPI_T interface to perform runtime tuning of MPI libraries. This plugin infrastructure is part of TAU, a popular performance analysis toolkit. While the plugin infrastructure is generic and can support any MPI implementation, we present a study of tuning applications running on MVAPICH, a widely used high-performance MPI implementation that offers a .wide range of performance metrics and tunable parameters through the MPI_T interface.

This poster introduces the Horizon 2020 READEX project (Runtime Exploitation of Application Dynamism for Energy-efficient eXascale computing) and its approach of dynamic application behavior analysis and tuning. The attention is paid to the manual application evaluation from the energy consumption point of view. This is the key step in exploring the possible gains of the runtime dynamic tuning. We present the effect of static (constant settings for the entire application runtime) and dynamic tuning (different settings for different parts of the application code) of the CPU core frequency, CPU uncore frequency and the number of active cores on the energy consumption of an HPC cluster running selected application. 
 
To show the potential dynamism we have evaluated basic kernels with various computational intensity (DGEMV, DGEMM and compute only kernel) and parallel I/O. Also, two full fledge applications: ESPRESO FEM library with FETI solvers and open-source CFD tool OpenFOAM were evaluated. The results show that these two complex applications can achieve energy savings of 21.3% and 17.7%, respectively, as a combination of static and dynamic tuning.