Show simple item record

dc.contributor.advisorDeng, Yuefan; Lindquist, Brenten_US
dc.contributor.authorGao, Yuxiangen_US
dc.contributor.otherDepartment of Applied Mathematics and Statisticsen_US
dc.description.abstractTask mapping has been explored intensively by computer and computational scientists for the purposes of maximizing utilization of computing resources by reducing communication and load imbalance. Task mapping is known to be NP-complete as a combinatorial optimization problem and, thus, many efforts intending to search for optimal solutions are not succeeding. All proposed mapping models to date represent parallel computers and applications as graphs with the networking capability abstracted as a supply matrix whose entries represent the inter-node communication costs and the communication requirements of the sub-tasks abstracted as a demand matrix whose entries represent the inter-subtask communication loads. The model is to minimize the objective function, defined as the hop-bytes metric. However, this hop-bytes metric is always written as a sum of the element-by-element products of the network supply and application demand matrices. This native and somewhat naive approach completely neglects the enormous symmetries in most networks and many applications and thus requires a huge number of unnecessary optimization steps. This thesis proposes a new formulation to investigate static task assignment. We develop a new formulation, by aid of eigen-analysis, to alleviate this very shortcoming of the past models. As expected, our formulation helps reduce the optimization steps in finding solution of the objective function. The extent of the speedup is not easily obtainable through analytical means but numerical experiments of mapping a suite of benchmarks, HPCC and NPB, onto 3D torus networks demonstrated that the simulated annealing process converges substantially faster with our new simulation than with the standard graph theory-based ones.en_US
dc.description.sponsorshipStony Brook University Libraries. SBU Graduate School in Department of Applied Mathematics and Statistics. Lawrence Martin (Dean of Graduate School).en_US
dc.formatElectronic Resourceen_US
dc.publisherThe Graduate School, Stony Brook University: Stony Brook, NY.en_US
dc.subject.lcshApplied Mathematicsen_US
dc.subject.othereigen analysis, simulated annealing, symmetry, task mappingen_US
dc.titleAn Algebraic Representation of the Task Mapping for Parallel Computingen_US
dc.description.advisorAdvisor(s): Yuefan Deng. Brent Lindquist. Committee Member(s): Xiangmin Jiao; Hong Qin.en_US
dc.embargo.period2 Yearsen_US

Files in this item


There are no files associated with this item.

This item appears in the following Collection(s)

Show simple item record