Cactus Application: Performance Predictions in Grid Environments

Abstract

Abstract. The Cactus software is representative for a whole class of scientific applications; typically those that are tightly coupled, have regular space decomposition, and huge memory and processor time requirements. Cactus proved to be a valuable tool for astrophysicists, who first initiated its development. However, today's fastest supercomputers are not powerful enough to perform realistically large astrophysics simulations with Cactus. The emergence of innovative resource environments like Grids satisfies this need for computational power. Our paper addresses issues related to the execution of applications like Cactus in Grid environments. We focus on two types of Grids: a set of geographically distributed supercomputers and a collection of the scale of one million Internet-connected workstations. We study the application performance on traditional systems, validate the theoretical results against experimental data, and predict performance in the two new environments.

A Performance Study for Iterative Stencil Loops on GPUs with Ghost Zone Optimizations

“…Ripeanu et al constructed a performance model that can predict the optimal ghost zone size [32], and they conclude the optimal ghost zone size is usually one in distributed environments. However, the performance model is based on message-passing and it does not model shared memory systems.…”

“…This ghost zone enlarges the tile with a perimeter overlapping neighboring tiles by multiple halo regions, as shown in Figure 1(b). The overlap allows each PE to generate its halo regions locally [32] for a number of iterations proportional to the size of the ghost zone. As Figure 1(b) demonstrates, ghost zones group loops into stages, where each stage operates on overlapping stacks of tiles, which we refer to as trapezoids.…”

“…Despite the ghost zone's potential benefit, an improper selection of the ghost zone size may negatively impact the overall performance. Previously, optimization techniques for ghost zones have only been proposed in message-passing based distributed environments [32]. These techniques no longer fit for modern chip multiprocessors (CMPs) for two reasons.…”

Supporting efficient execution in heterogeneous distributed computing environments with cactus and globus

“…We have built a model to evaluate the expected performance in a Grid environment [21]. The model is parameterized with execution environment data (number and performance of processors, network performance) and application data (problem size, ghostzone size, etc.…”

Near-optimal adaptive control of a large grid application

“…For example, Ripeanu et al [21] compute the execution time for a load-balanced finite difference application as a function of (a) the data assigned to each node, (b) redundant work computed by other nodes to reduce communication between nodes, and (c) the communication time. These calculations are used at runtime to select the amount of redundant work per node for the measured Grid communication costs.…”