Performance models for the processor farm paradigm

Abstract

In this paper, we describe the design, implementation, and modeling of a runtime kernel to support the processor farm paradigm on multicomputers. We present a general topology-independent framework for obtaining performance models to predict the performance of the start-up, steady-state, and wind-down phases of a processor farm. An algorithm is described, which for any interconnection network determines a tree-structured subnetwork that optimizes farm performance. The analysis technique is applied to the important case of k-ary tree topologies. The models are compared with the measured performance on a variety of topologies using both constant and varied task sizes.

Constructing generic data-farm templates

“…[45] is an elaboration of this model and also contains a thorough review of data farms, making the link with skeletons.…”

Cross-Platform Performance Prediction of Parallel Applications Using Partial Execution

“…Parallel Program Performance Prediction: There have been numerous previous studies of performance prediction for parallel programs. Many of these studies are built upon performance modeling techniques (e.g., [1,9,13,18,24,31,34]), requiring either in-depth knowledge of the applications to build analytical models (e.g., [2,21,28,36,37]) or special compiler/instrumentation tools to infer such knowledge from parallel codes (e.g., [6,8,15,24]). With careful modeling of applications and platforms, many of these previous studies achieved high prediction accuracy.…”

“…In this realm, formal performance modeling and simulation have played a central role [1,2,6,8,9,13,15,18,21,24,28,31,34,36,37]. While model-based prediction has become indispensable for future architectures, its accuracy is increasingly compromised for existing architectures due to their complexity and heterogeneity.…”

Providing structure for medium-grained distributed object-based computation

“…Feedback or feedforward loops enable synchronous constraints to be introduced but for practicality only one such loop may occur at each input site. The PPF pattern is an extension and generalisation of a number of data-farming schemes: some NOWbased [31]; some based on dedicated multicomputers [38]; some constructed as algorithmic skeletons [28]; and some implicitly invoked in parallel extensions of C++ [20]. …”