6. Partitioning and Load Balancing for Emerging Parallel Applications and Architectures

Abstract

Chapter 1. Partitioning and Load Balancing several different algorithms enable developers to easily compare methods to determine their effectiveness in applications [24, 26, 60]. Prior efforts have focused primarily on partitioning for homogeneous computing systems, where computing power and communication costs are roughly uniform. Wider acceptance of parallel computing has lead to an explosion of new parallel applications. Electronic circuit simulations, linear programming, materials modeling, crash simulations, and data mining are all adopting parallel computing to solve larger problems in less time. And the parallel architectures they use have evolved far from uniform arrays of multiprocessors. While homogeneous, dedicated parallel computers can offer the highest performance, their cost often is prohibitive. Instead, parallel computing is done on everything from networks of workstations to clusters of shared-memory processors to grid computers. These new applications and architectures have reached the limit of standard partitioners' effectiveness; they are driving development of new algorithms and software for partitioning. This paper surveys current research in partitioning and dynamic load balancing, with special emphasis on work presented at the 2004 SIAM Conference on Parallel Processing for Scientific Computing. "Traditional" load-balancing methods are summarized in §1.2. In §1.3, we describe several non-traditional applications along with effective partitioning strategies for them. Some non-traditional approaches to load balancing are described in §1.4. In §1.5, we describe partitioning goals that reach beyond typical load-balancing objectives. And in §1.6, we address load-balancing issues for non-traditional architectures.

Combinatorial Problems in Solving Linear Systems

“…As this is the subject of another survey [23], we do not cover this issue. Instead, we refer the reader to [37,49,110,188] for this interaction and to [71] for different aspects and a survey of earlier studies on parallelization.…”

Platform Calibration for Load Balancing of Large Simulations: TLM Case

“…This graph is matched with the graph generated for the application. In [11] is described a resource-aware partitioning where information about a computing environment is combined with traditional partitioning algorithms. The approach collects information about the computing environment and processes it for partitioning use.…”

MCD: Mesh Closure Detection for Localized Load Balancing in Scientific Applications

“…A wealth of partitioning research exists for mesh-based partial differential equation (PDE) solvers (e.g., finite volume and finite element methods) and their sparse linear solvers [7]. Conceptually geometric methods have proven to be highly effective for particle simulations, while providing reasonably good decompositions for mesh-based solvers.…”

“…They only use objects' weights and physical coordinates, and assign equal object weight to processors while grouping physically close objects within subdomains [7]. Recursive…”