Search

Rapid growth in data, computational methods, and computing power is driving a remarkable revolution in what variously is termed machine learning (ML), statistical learning, computational learning, and artificial intelligence. In addition to highly visible successes in machine-based natural language translation, playing the game Go, and self-driving cars, these new technologies also have profound implications for computational and experimental science and engineering, as well as for the exascale computing systems that the Department of Energy (DOE) is developing to support those disciplines. Not only do these learning technologies open up exciting opportunities for scientific discovery on exascale systems, they also appear poised to have important implications for the design and use of exascale computers themselves, including high-performance computing (HPC) for ML and ML for HPC. The overarching goal of the ExaLearn co-design project is to provide exascale ML software for use by Exascale Computing Project (ECP) applications, other ECP co-design centers, and DOE experimental facilities and leadership class computing facilities.

We discuss algorithm-based resilience to silent data corruptions (SDCs) in a task-based domain-decomposition preconditioner for partial differential equations (PDEs). The algorithm exploits a reformulation of the PDE as a sampling problem, followed by a solution update through data manipulation that is resilient to SDCs. The implementation is based on a server-client model where all state information is held by the servers, while clients are designed solely as computational units. Scalability tests run up to ∼51 K cores show a parallel efficiency greater than 90%. We use a 2D elliptic PDE and a fault model based on random single and double bit-flip to demonstrate the resilience of the application to synthetically injected SDC. We discuss two fault scenarios: one based on the corruption of all data of a target task, and the other involving the corruption of a single data point. We show that for our application, given the test problem considered, a four-fold increase in the number of faults only yields a 2% change in the overhead to overcome their presence, from 7% to 9%. We then discuss potential savings in energy consumption via dynamic voltage/frequency scaling, and its interplay with fault-rates, and application overhead.

We present a methodology to assess the predictive fidelity of multiscale simulations by incorporating uncertainty in the information exchanged between the components of an atomistic-to-continuum simulation. We account for both the uncertainty due to finite sampling in molecular dynamics (MD) simulations and the uncertainty in the physical parameters of the model. Using Bayesian inference, we represent the expensive atomistic component by a surrogate model that relates the long-term output of the atomistic simulation to its uncertain inputs. We then present algorithms to solve for the variables exchanged across the atomistic-continuum interface in terms of polynomial chaos expansions (PCEs). We consider a simple Couette flow where velocities are exchanged between the atomistic and continuum components, while accounting for uncertainty in the atomistic model parameters and the continuum boundary conditions. Results show convergence of the coupling algorithm at a reasonable number of iterations. The uncertainty in the obtained variables significantly depends on the amount of data sampled from the MD simulations and on the width of the time averaging window used in the MD simulations.

In this paper, a series of algorithms are proposed to address the problems in the NASA Langley Research Center Multidisciplinary Uncertainty Quantification Challenge. A Bayesian approach is employed to characterize and calibrate the epistemic parameters based on the available data, whereas a variance-based global sensitivity analysis is used to rank the epistemic and aleatory model parameters. A nested sampling of the aleatory–epistemic space is proposed to propagate uncertainties from model parameters to output quantities of interest.

The move towards extreme-scale computing platforms challenges scientific simulations in many ways. Given the recent tendencies in computer architecture development, one needs to reformulate legacy codes in order to cope with large amounts of communication, system faults, and requirements of low-memory usage per core. In this work, we develop a novel framework for solving PDEs via domain decomposition that reformulates the solution as a state of knowledge with a probabilistic interpretation. Such reformulation allows resiliency with respect to potential faults without having to apply fault detection, avoids unnecessary communication, and is generally well-suited for rigorous uncertainty quantification studies that target improvements of predictive fidelity of scientific models. We demonstrate our algorithm for one-dimensional PDE examples where artificial faults have been implemented as bit flips in the binary representation of subdomain solutions.

We present a task-based domain-decomposition preconditioner for partial differential equations (PDEs) resilient to silent data corruption (SDC) and hard faults.The algorithm exploits a reformulation of the PDE as a sampling problem, followed by a regression-based solution update that is resilient to SDC. We adopt a server-client model implemented using the User Level Fault Mitigation MPI (MPI-ULFM). All state information is held by the servers, while clients only serve as computational units. The task-based nature of the algorithm and the capabilities of ULFM are complemented at the algorithm level to support missing tasks, making the application resilient to hard faults affecting the clients.Weak and strong scaling tests up to ~115k cores show an excellent performance of the application with efficiencies above 90%, demonstrating the suitability to run at large scale. We demonstrate the resilience of the application for a 2D elliptic PDE by injecting SDC using a random single bit-flip model, and hard faults in the form of clients crashing. We show that in all cases, the application converges to the right solution. We analyze the overhead caused by the faults, and show that, for the test problem considered, the overhead incurred due to SDC is minimal compared to that from the hard faults.

The need to make large-scale scientific simulations resilient to the shrinking and growing of compute resources arises from exascale computing and adverse operating conditions (fault tolerance). It can also arise from the cloudcomputing context where the cost of these resources can fluctuate. In this paper, we describe how the Sparse Grid Combination Technique can make such applications resilient to shrinking compute resources. The solution of the non-trivial issues of dealing with data redistribution and on-the-fly malleability of process grid information and ULFM MPI communicatorsare described. Results on a 2D advection solver indicate that process recovery time is significantly reduced from the alternate strategy where failed resources are replaced, overall execution time is actually improved from this case and for checkpointing and the execution error remains small, even when multiple failures occur.

This work focuses on quantifying the effect of intrinsic (thermal) noise and parametric uncertainty in molecular dynamics (MD) simulations. We consider isothermal, isobaric MD simulations of TIP4P (or four-site) water at ambient conditions, $T=298$ K and $P=1$ atm. Parametric uncertainty is assumed to originate from three force-field parameters that are parametrized in terms of standard uniform random variables. The thermal fluctuations inherent in MD simulations combine with parametric uncertainty to yield nondeterministic, noisy MD predictions of bulk water properties. Relying on polynomial chaos (PC) expansions, we develop a framework that enables us to isolate the impact of parametric uncertainty on the MD predictions and control the effect of the intrinsic noise. We construct the PC representations of quantities of interest (QoIs) using two different approaches: nonintrusive spectral projection (NISP) and Bayesian inference. We verify a priori the legitimacy of the NISP approach by verifying that the...

Micro- and nanofluidics pose a series of significant challenges for science-based modeling. Key among those are the wide separation of length- and timescales between interface phenomena and bulk flow and the spatially heterogeneous solution properties near solid-liquid interfaces. It is not uncommon for characteristic scales in these systems to span nine orders of magnitude from the atomic motions in particle dynamics up to evolution of mass transport at the macroscale level, making explicit particle models intractable for all but the simplest systems. Recently, atomistic-to-continuum (A2C) multiscale simulations have gained a lot of interest as an approach to rigorously handle particle-level dynamics while also tracking evolution of large-scale macroscale behavior. While these methods are clearly not applicable to all classes of simulations, they are finding traction in systems in which tight-binding, and physically important, dynamics at system interfaces have complex effects on the slower-evolving large-scale evolution of the surrounding medium. These conditions allow decomposition of the simulation into discrete domains, either spatially or temporally. In this paper, we describe how features of domain decomposed simulation systems can be harnessed to yield flexible and efficient software for multiscale simulations of electric field-driven micro- and nanofluidics.

We present a resilient domain-decomposition preconditioner for partial differential equations (PDEs). The algorithm reformulates the PDE as a sampling problem, followed by a solution update through data manipulation that is resilient to both soft and hard faults. We discuss an implementation based on a server-client model where all state information is held by the servers, while clients are designed solely as computational units. Servers are assumed to be “sandboxed”, while no assumption is made on the reliability of the clients. We explore the scalability of the algorithm up to \(\sim \)12k cores, build an SST/macro skeleton to extrapolate to \(\sim \)50k cores, and show the resilience under simulated hard and soft faults for a 2D linear Poisson equation.

The NASA Langley Multidisciplinary Uncertainty Quantification Challenge1 presents several questions, formulated around computer models that describe realistic aeronautical applications. These models are presented in a “blackbox” formulation to encourage discipline-independent approaches. The schematic in Fig. 1 illustrates the general structure of the model inputs, outputs, and quantities of interest.

The UQ Toolkit (UQTk) is a collection of libraries and tools for the quanti cation of uncer- tainty in numerical model predictions. Version 2.0 o ers intrusive and non-intrusive methods for propagating input uncertainties through computational models, tools for sensitivity anal- ysis, methods for sparse surrogate construction, and Bayesian inference tools for inferring parameters from experimental data. This manual discusses the download and installation process for UQTk, provides pointers to the UQ methods used in the toolkit, and describes some of the examples provided with the toolkit.

Simulation of high-performance computing (HPC) systems plays a critical role in their development - especially as HPC moves toward the co-design model used for embedded systems, tying hardware and software into a unified design cycle. Exploring system-wide tradeoffs in hardware, middleware and applications using high-fidelity cycle-accurate simulation, however, is far too costly. Coarse-grained methods can provide efficient, accurate simulation but require rigorous uncertainty quantification (UQ) before using results to support design decisions. We present here SST/macro, a coarse-grained structural simulator providing flexible congestion models for low-cost simulation. We explore the accuracy limits of coarse-grained simulation by deriving error distributions of model parameters using Bayesian inference. Propagating these uncertainties through the model, we demonstrate SST/macro's utility in making conclusions about performance tradeoffs for a series of MPI collectives. Low-cost and high-accuracy simulations coupled with UQ methodology make SST/macro a powerful tool for rapidly prototyping systems to aid extreme-scale HPC co-design.

Computational frameworks can significantly assist in the construction, extension and maintenance of simulation codes. As the nature of problems addressed by computational means has grown in complexity, such frameworks have evolved to incorporate a commensurate degree of sophistication, both in terms of the numerical algorithms that they accommodate as well as the software architectural discipline they impose on their users. In this chapter, we discuss a component framework, the Common Component Architecture (CCA), for developing scientific software, and describe how it has been used to develop a toolkit for simulating reacting flows. In particular, we will discuss why a component architecture was chosen and the philosophy behind the particular software design. Using statistics drawn from the toolkit, we will analyze the code structure and investigate to what degree the aims of the software design were actually realized. We will explore how CCA was employed to design a high-order simulation code on block-structured adaptive meshes, as well as a simulation capacity for adaptive stiffness reduction in detailed chemical models. We conclude the chapter with two reacting flow studies performed using the above-mentioned computational capabilities.