Estimation of execution parameters takes centre stage in automatic offloading of complex biomedical workflows to cloud and high performance facilities. Since ordinary users have no or very limited knowledge of the performance characteristics of particular tasks in the workflow, the scheduling system has to have the capabilities to select appropriate amount of compute resources, e.g., compute nodes, GPUs, or processor cores and estimate the execution time and cost. The presented approach considers a fixed set of executables that can be used to create custom workflows, and collects performance data of successfully computed tasks. Since the workflows may differ in the structure and size of the input data, the execution parameters can only be obtained by searching the performance database and interpolating between similar tasks. This paper shows it is possible to predict the execution time and cost with a high confidence. If the task parameters are found in the performance database, the mean interpolation error stays below 2.29%. If only similar tasks are found, the mean interpolation error may grow up to 15%. Nevertheless, this is still an acceptable error since the cluster performance may vary on order of percent as well.
In a simple undirected graph, we introduce a special connectedness induced by a set of paths of length 2. We focus on the 8-adjacency graph (with the vertex set ℤ^2 ) and study the connectedness induced by a certain set of paths of length 2 in the graph. For this connectedness, we prove a digital Jordan curve theorem by determining the Jordan curves, i.e., the circles in the graph that separate ℤ^2 into exactly two connected components. These Jordan curves are shown to have an advantage over those given by the Khalimsky topology on ℤ^2 .
The rapid development of novel procedures in medical ultrasonics, including treatment planning in therapeutic ultrasound and image reconstruction in photoacoustic tomography, leads to increasing demand for large-scale ultrasound simulations. However, routine execution of such simulations using traditional methods, e.g., finite difference time domain, is expensive and often considered intractable due to the computational and memory requirements. The k-space corrected pseudospectral time domain method used by the k-Wave toolbox allows for significant reductions in spatial and temporal grid resolution. These improvements are achieved at the cost of all-to-all communication, which are inherent to the multi-dimensional fast Fourier transforms. To improve data locality, reduce communication and allow efficient use of accelerators, we recently implemented a domain decomposition technique based on a local Fourier basis. In this paper, we investigate whether it is feasible to run the distributed k-Wave implementation on the Salomon cluster equipped with 864 Intel Xeon Phi (Knight’s Corner) accelerators. The results show the immaturity of the KNC platform with issues ranging from limited support of Infiniband and LustreFS in Intel MPI on this platform to poor performance of 3D FFTs achieved by Intel MKL on the KNC architecture. Yet, we show that it is possible to achieve strong and weak scaling comparable to CPU-only platforms albeit with the runtime 1.8× to 4.3× longer. However, the accounting policy for Salomon’s accelerators is far more favorable and thus their employment reduces the computational cost significantly.
The behavior of a parallel application can be presented in many ways, but performance visualization tools usually focus on communication graphs and runtime of processes or threads in specific (groups of) functions. A different approach is required when searching for the optimal configuration of tunable parameters, for which it is necessary to run the application several times and compare the resource consumption of these runs. We present RADAR visualizer, a tool that was originally developed to analyze such measurements and to detect the optimal configuration for each instrumented part of the code. In this case, the optimum was defined as the minimum energy consumption of the whole application, but any other metric can be defined. RADAR visualizer presents the application behavior in several graphical representations and tables including the amount of savings that can be reached. Together with our MERIC library, we provide a complete toolchain for HPC application behavior monitoring, data analysis, and graphical representation. The final part is performing dynamic tuning (applying optimal settings for each region during the application runtime) for the production runs of the analyzed application.
It is expected that with the appearance of exascale supercomputers the mean time between failure in supercomputers will decrease. Classical checkpoint-restart approaches are too expensive at that scale. Local-failure local-recovery (LFLR) strategies are an option that promises to leverage the costs, but actually implementing it into any sufficiently large simulation environment is a challenging task. In this paper we discuss how LFLR methods can be incorporated in a PDE framework, focussing at the linear solvers as the innermost component. We discuss how Krylov solvers can be modified to support LFLR, and present numerical tests. We exemplify our approach by reporting on the implementation of these features in the Dune framework, present C++ software abstractions, which simplify the incorporation of LFLR techniques and show how we use these in our solver library. To reduce the memory costs of full remote backups, we further investigate the benefits of lossy compression and in-memory checkpointing.
