Identifier

etd-11172014-141748

Degree

Master of Science (MS)

Department

Electrical and Computer Engineering

Document Type

Thesis

Abstract

High Performance Computation (HPC) requires a proper and efficient scheme for distribution of the computational workload across different computational nodes. The HPX (High Performance ParalleX) runtime system currently lacks a module that automates data distribution process so that the programmer does not have to manually perform data distribution. Further, there is no mechanism allowing to perform load balancing of computations. This thesis addresses that issue by designing and developing a user friendly programming interface conforming to the C++11/14 Standards and integrated with HPX which enables to specify various distribution parameters for a distributed vector. We present the three different distribution policies implemented so far: block, cyclic, and block-cyclic. These policies influence the way the distributed vector maps any global (linear) index into the vector onto a pair of values describing the number of the (possibly remote data partition) and the corresponding local index. We present performance analysis results from applying the different distribution policies to calculating the Mandelbrot set; an example of an ‘embarrassingly parallel’ computation. For this benchmark we use an instance of a distributed vector where each element holds a tuple for the current index and the value of related to an individual pixel of the generated Mandelbrot plot. We compare the influence of different distribution policies and their corresponding parameters on the overall execution time of the calculation. We demonstrate that the block-cyclic distribution policy yields best results for calculating the Mandelbrot set as it more evenly load balances the computation across the computational nodes. The provided API and implementation gives the user a high level an abstraction for developing applications while hiding low-level data distribution details.

Date

2014

Document Availability at the Time of Submission

Release the entire work immediately for access worldwide.

Committee Chair

Kaiser, Hartmut

DOI

10.31390/gradschool_theses.2169

Download

Share

COinS