[LA-SiGMA-TESC] Technologies for Extreme Scale Computing seminar on Thursday (8/28)

[Wei P Feinstein]

Hi all,


Welcome back. Our weekly seminar for the fall will start this week on Thursday (8/28) at 12:00pm-1:00pm in 1034 Digital Media Center Theatre.


Roozbeh Karimi, a graduate student in the department of computer science and engineering, is going to present his research titled "Exploiting Sparse Matrix Structure in an Arnoldi Iteration for a Quantum Monte-Carlo simulation on Many-core architectures".

Abstract:
Sparse matrix-vector multiplication (SpMV) is one of the most time consuming parts of Quantum Monte-Carlo simulations. The matrix we're interested in is 1024-by-1024, and has 3574 nonzero elements. For our matrix, conventional Sparse matrix formats like the COO, CSR, ELL and DIA require 56kiB, 46kiB, 84kiB and 328kiB respectively; also, due to irregular memory access patterns, the more efficient (memory-wise) COO and CSR formats preclude peak execution rates on existing SIMD architectures like GPU and Xeon Phi. We propose a especially tailored scheme to exploit the structure of our sparse matrix. Our method reduces the memory requirements for the sparse matrix down to 9kiB, allowing us to keep everything during the SpMV operation within the L1 cache of the Xeon Phi. Also, we propose taking advantage of the instruction cache to further reduce cache requirements; this also allows us to utilize compiler optimizations to increase performance. Another optimization we have proposed is to use a skewed storage format to reduce the gather/scatter instructions associated with SpMV.


PIZZA AND DRINK WILL BE PROVIDED.

Presentation time: 12:00PM~1:00PM CST
Location: 1034 Digital Media Center Theatre


The meeting will be shared through adobe connect at http://connect.lsu.edu/gpu/


Best,
Wei

Wei Feinstein, Ph.D.
Center for Computation and Technology
Computational Biology
Louisiana State University
E-mail: wfeinstein at lsu.edu
Lab: http://brylinski.cct.lsu.edu/
Office: 401 Choppin Hall, 2060 DMC
Tel: 251-599-3814
--------------------------------------------------
LASiGMA-gpu mailing list
LASiGMA-gpu at loni.org<mailto:LASiGMA-gpu at loni.org>
https://mail.loni.org/mailman/listinfo/lasigma-gpu
wiki: https://wiki.cct.lsu.edu/la-sigma/
http://www.institute.loni.org/lasigma/lasigma-gpu.php

-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://mail.loni.org/mailman/private/lasigma-gpu/attachments/20140826/09a7fab1/attachment-0001.html