Predicting the Cost and Benefit of Adapting Data Parallel Applications in Clusters

Abstract

This paper examines the problem of adapting data parallel applications in a shared dynamic environment of PC or workstation clusters. We developed an analytic framework to compare and contrast a wide range of adaptation strategies: dynamic load balancing, migration, processor addition and removal. These strategies have been evaluated with respect to the cost and benefit they provide for three representative parallel applications: an iterative jacobi solver for Laplace's equation, gaussian elimination with partial pivoting, and a gene sequence comparison application. We found that the cost and benefit of each method can be predicted with high accuracy (within 10%) for all applications and show that the framework is applicable to a wide variety of parallel applications. We then show that accurate prediction allows the most appropriate method to be selected dynamically. Performance improvement for the three applications ranged from 25% to 45% using our adaptation library. In addition, we dispel the conventional wisdom that migration is too expensive, and show that it can be beneficial even for running parallel applications with non-trivial communication. # 2002 Elsevier Science (USA)

A Survey of Recent Developments in Rotorcraft Design Optimization

“…201 Finally, as parallel computing becomes easier with personal computer or workstation cluster computing, the ability of stochastic optimization methods for solving helicopter optimization problems on personal computer clusters should be explored. 185 An issue that would greatly help the helicopter optimization community is the presence of test problems. Because the analysis codes used differ for different problems, one approach can be to create an approximate problem using surrogate functions from an established aeroelastic analysis and present it as a test problem.…”

Adaptive middleware supporting scalable performance for high-end network services

“…In order to achieve accurate predictions of the application run-time, many conventional approaches used in scheduling research (Maheswaran and Siegel, 1998;Subramani et al, 2002;Takefusa et al, 2001;Vadhiyar and Dongarra, 2002;Weissman, 2002) assume that accurate performance models are available that include low-level details of the applications (e.g., operation counts, message size). Kapadia et al (1999) use three instance-based learning algorithms to predict resource usage for each application run.…”

Load balancing and optimization of distributed proximity effect correction on a temporally heterogeneous cluster

“…In [15], a "pre-computation-based" dynamic load balancing on heterogeneous systems was described, but only local balancing was allowed. In [16], a prediction method for dynamic load balancing was described. However, the possible dependency of the optimal threshold on the task size was not considered.…”