Type of Document Dissertation Author Amatya, Vinay Chandra URN etd-11172014-122205 Title Parallel Processes in HPX: Designing an Infrastructure for Adaptive Resource Management Degree Doctor of Philosophy (Ph.D.) Department Computer Science Advisory Committee
Advisor Name Title Kaiser, Hartmut Committee Chair Karki, Bijay B. Committee Co-Chair Busch, Konstantin Committee Member Ramanujam, J. Committee Member Stoltzfus, Neal Dean's Representative Keywords
- extreme scale computing
- resource management
- load balancing
- concurrency management
Date of Defense 2014-11-10 Availability unrestricted AbstractAdvancement in cutting edge technologies have enabled better energy efficiency as well
as scaling computational power for the latest High Performance Computing(HPC) systems.
However, complexity, due to hybrid architectures as well as emerging classes
of applications, have shown poor computational scalability using conventional execution models.
Thus alternative means of computation, that addresses the bottlenecks in computation, is warranted.
More precisely, dynamic adaptive resource management feature, both from systems as well as
application's perspective, is essential for better computational scalability and efficiency.
This research presents and expands the notion of Parallel Processes as a placeholder for
procedure definitions, targeted at one or more synchronous domains, meta data for
computation and resource management as well as infrastructure for dynamic
policy deployment. In addition to this, the research presents additional guidelines for a
framework for resource management in HPX runtime system.
Further, this research also lists design principles for scalability of Active
Global Address Space (AGAS), a necessary feature for Parallel Processes. Also,
to verify the usefulness of Parallel Processes, a
preliminary performance evaluation of different task scheduling policies is carried out
using two different applications. The applications used are: Unbalanced Tree Search,
a reference dynamic graph application, implemented by this research in HPX and MiniGhost,
a reference stencil based application using bulk synchronous parallel model. The
results show that different scheduling policies provide better performance for different
classes of applications; and for the same application class, in certain instances,
one policy fared better than the others, while vice versa in other instances, hence supporting
the hypothesis of the need of dynamic adaptive resource management infrastructure,
for deploying different policies and task granularities, for scalable distributed computing.
Filename Size Approximate Download Time (Hours:Minutes:Seconds)
28.8 Modem 56K Modem ISDN (64 Kb) ISDN (128 Kb) Higher-speed Access amatya_dissertation_1.pdf 5.60 Mb 00:25:54 00:13:19 00:11:39 00:05:49 00:00:29
If you have questions or technical problems, please Contact LSU-ETD Support.