Page History

Luc Jaulmes

Checkpointless exact recovery techniques for Krylov-based iterative methods

Lokman Rahmani

Lucas Nussbaum

Evaluating exascale HPC runtimes through emulation with Distem

Sanjay Kale

Temperature, Power and Energy: How an Adaptive Runtime can optimize them.

Jonathan Jenkins

Towards Simulating Extreme-scale Distributed Systems

 Simulating future extreme-scale parallel/distributed systems can be an important component in understanding these systems at a scale at which prototyping cannot feasibly reach. For HPC, big-data/cloud, or other computing/analysis platforms, the design decisions for developing systems that scale beyond current-generation systems are multi-dimensional in nature. For example, these decisions encompass distributed storage software/hardware solutions, network topologies within and between computing centers, algorithms for data analysis and compute services in heterogeneous software/hardware environments, etc., each of which can potentially be rich targets for exploring via a simulation-based approach. This talk will examine our ongoing work in developing a simulation model framework using parallel discrete event simulation to examine various design aspects of extreme-scale distributed systems. As an exemplar, simulation of protocols used in distributed storage systems will be examined in detail.

Timothy Armstrong

Towards Dynamic Dataflow Composition for Extreme-Scale Applications with Heterogeneous Tasks
Parallel applications are increasingly built from heterogeneous software components that use diverse programming models, such as message-passing, threads, CUDA, and OpenCL on heterogeneous hardware resources such as CPUs and GPUS.  Getting these components to interoperate is a challenge in itself, which is further complicated by complex cross-cutting concerns such as scheduling, overlapping of communication and computation, fault-tolerance, and energy efficiency. Parallel execution models offer the hope of making these challenges more managable for application programmers by unifying heterogeneous components into a more uniform framework.  One such model is data-driven task parallelism, in which massive numbers of tasks are dynamically assigned to compute resources and communication and synchronization is based on explicit data dependencies.  Swift is a high-level scripting
language that provides a simple yet powerful way of expressing data-driven task parallelism.  This talk discusses our current progress and future challenges on a compiler and runtime system that allows Swift to scale to hundreds of thousands of cores.

Juan González

Performance Analytics: Understanding Parallel Applications using Cluster Analysis and Sequence Analysis.
Due to the increasing complexity of HPC systems and applications it is strictly necessary to maximize the insight of the performance data extracted from an application execution. This is the mission of the Performance Analytics field. In this talk we introduce two Performance Analytics techniques. First, we demonstrate how it is possible to capture the computation structure of parallel applications at fine grain by using density-based cluster algorithms. Second, we introduce the use of multiple sequence alignment algorithms to asses the quality of this computation structure."

An alternate interpretation of the Full Approximation Scheme (FAS) multigrid method creates relationships between levels that can be exploited to eliminate communication on fine grids, avoid storage of fine grids, avoid "visiting" fine grids away from active nonlinearities, accelerate recomputation from checkpoints, and use fine-to-coarse compatibility to check for silent data corruption in fine grid state. This talk will present the algorithmic structure, new results with ultra-low-communication parallel multigrid, and directions for future research.

Reducing Complexity in Algebraic Solvers

The coarse-level sparse matrices operations are defined through the Galerkin product, R A P — i.e., restriction, operator, and interpolation.  Consequently, we look at two methods that reduce this complexity: an approach that filters P  and a method that builds a coarse level through a non-Galerkin construction.  To this end we first introduce a root-node based approach to multigrid, which  can be viewed as a hybrid of classical and aggregation based multigrid methods.  We give an overview and show how the complexity and convergence of the multigrid cycle can be controlled through selective filtering in a root-node setting.  In addition, we look at a non-Galerkin algebraic framework where we are able to model the performance and note the performance gains in selectively filtering coarse-level operators.

Vincent Baudoui

Round-off error propagation in large-scale applications

Round-off errors coming from numerical calculation finite precision can lead to catastrophic losses in significant numbers when they accumulate. They will become more and more overriding in the future as the problem size increases with the refinement of numerical simulations. Existing analytical bounds for round-off errors are known to be poorly scalable and they become quite useless for large problems. That is why the propagation of round-off errors throughout a computation needs to be better understood in order to ensure large-scale application results accuracy. We study here a round-off error estimation method based on first order derivatives computed thanks to algorithmic differentiation techniques. It can help following the error propagation through a computational graph and identifying the sensitive sections of a code. It has been experimented on well known LU decomposition algorithms that are widely used to solve linear systems. We will present some examples as well as challenges that need to be tackled as part of future research work in order to set up a strategy to analyze round-off error propagation in large-scale problems.