Home > Events > Brian Page - Oral Candidacy

Brian Page - Oral Candidacy

Start: 12/7/2018 at 8:00AM
End: 12/7/2018 at 10:30AM
Location: 165 Fitzpatrick Hall
Attendees: Faculty and students are welcome to attend the presentation portion of the defense.
Add to calendar:
iCal vCal

Brian Page

Oral Candidacy

December 7, 2018     117I Cushing Hall      8:00 am

Adviser:  Dr. Peter Kogge

Committee Members;

Dr. Sharon Hu      Dr. Christian Poellabauer      Dr. Douglas Thain

 

Title:

Strong Scalability of Hyper-Sparse Problems

 

Abstract:


The ever increasing size of large data sets presents complex computational challenges for analysts and researchers alike. Nearly every field of engineering deals with problems resulting in the generation of sparse data sets, often in the form of sparse data types. Sparse data types posses irregular access patterns due to the data compression techniques intended to boost performance by eliminating the need to store superfluous data. While there has been much work focused on weak scaling of sparse problems, there has been relatively little research into the strong scaling of sparse problems. While both weak and strong scaling have advantages and disadvantages, as data sets continue to increase the need for efficient strong scaling of sparse problems is rapidly becoming apparent. 

We intend to investigate the performance characteristics of sparse problems in an effort to identify elements of implementation which impact performance, as well as to what degree they do so. First, I propose the development of separate hybrid SpMV codes targets to GPUs and the Emu migrating thread architecture. Second a hybrid bipartite matching application will be created using the vertex-centric HavoqGT framework. Third a code base of generalized sparse operators intended for strong will be developed and evaluated. Lastly a predictive hybrid sparse model will encompass the knowledge obtained during the preceding studies. This proposal seeks to broaden understanding into the strong scaling potential of irregular sparse problems, aiding in the push for exascale technologies.