Abstract: Modern asynchronous runtime systems allow the re-thinking of large-scale scientific applications. With the example of a simulator of morphologically detailed neural networks, we show how detaching from the commonly used bulk-synchronous parallel (BSP) execution allows for the increase of prefetching capabilities, better cache locality, and a overlap of computation and communication, consequently leading to a lower time to solution. Our strategy removes the operation of collective synchronization of ODEs’ coupling information, and takes advantage of the pairwise time dependency between equations, leading to a fully-asynchronous exhaustive yet not speculative stepping model. Combined with fully linear data structures, communication reduce at compute node level, and an earliest equation steps first scheduler, we perform an acceleration at the cache level that reduces communication and time to solution by maximizing the number of timesteps taken per neuron at each iteration. Our methods were implemented on the core kernel of the NEURON scientific application. Asynchronicity and distributed memory space are provided by the HPX runtime system for the ParalleX execution model. Benchmark results demonstrate a superlinear speedup that leads to a reduced runtime compared to the bulk synchronous execution, yielding a speedup between 25% to 65% across different compute architectures, and in the order of 15% to 40% for distributed executions.

Abstract: It is widely known that the worldwide development of civilization diseases (especially in the second half of the twentieth century) is the cause of the increase in mortality not caused by death from natural causes. In Poland, the most common causes of death, both for women and men, include cancer and cardiovascular disease. The aim of the article is to propose a method of modeling the life expectancy index based on the non-Gaussian linear scalar filter model stand on death rates after eliminating one or both of the above causes of death. The obtained results indicate that their elimination may be expected to extend life expectancy by several or more years depending on the cause of death and gender.

Abstract: Ordinary differential equations are widely used for the mathematical modeling of complex systems in biology and statistics. Since the analysis of such models needs to be performed using numerical integration, many applications can be gravely limited by the computational cost. This paper present a general-purpose integrator that runs massively parallel on graphics processing units. By minimizing thread divergence and bundling similar tasks using linear regression, execution time can be reduced by 40-80% when compared to a naive GPU implementation. Compared to a 36-core CPU implementation, a 150 fold runtime improvement is measured.

Abstract: Understanding the evolution of atomistic systems is essential in various fields such as materials science, biology, and chemistry. The gold standard for these calculations is molecular dynamics, which simulates the dynamical interaction between pairs of molecules. The main challenge of such simulation is the numerical complexity, given a vast number of atoms over a long time scale. Furthermore, such systems often contain exponentially many optimal states, and the simulation tends to get trapped in local configurations. Recent developments leverage the existing temporal evolution of the system to improve the stability and scalability of the method; however, they suffer from large data storage requirements. To efficiently compress the data while retaining the basins of attraction, we have developed a framework to determine the acceptable level of compression for an optimization method by application of a Kantorovich-type theorem, using binary digit rounding as our compression technique. Choosing the Lennard-Jones potential function as a model problem, we present a method for determining the local Lipschitz constant of the Hessian with low computational cost, thus allowing the use of our technique in real-time computation.

Abstract: Hydraulic Tomography (HT) has become one of the most sophisticated methods to characterize aquifer heterogeneity and in some experiments it has proved to be an accurate technique, but in order to achieve this goal, it is needed to perform several pumping/injection tests and to have enough measurements at each test. Also, during the solution of the inverse problem, the groundwater flow equation is solved numerically many times and thus the computational time can be very large, specially when a 3D or a transient models is used. In this work we present a new approach to model the aquifer heterogeneity based in a Gaussian Mixture Model, the proposed approach improves computation time and accuracy of the HT experiment and also it tries address the problems involved in the inverse problem, as is the effect of noisy data, the need of many pumping/injection tests and the lack of resolution when the distribution of the aquifer conductivity does not correspond to a Gaussian distribution. In synthetic experiments this approach was able to achieve one fifth of the error in the estimation of the conductivity field than one of the most used inversion methods for HT (SSLE/VSAFT), in one fourth of the computation time. In a steady-state sandbox experiment, it detected the main layers of conductivity in one fourth of the computational time than VSAFT, including a layer that was only detected with VSAFT when a transient model was used to perform the HT.