cwpearson/hugo-cwpearson
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

SIAM PP talk and abstract

Browse Source
This commit is contained in:
Carl Pearson
2022-02-22 14:48:41 -07:00
parent 20bffa7937
commit 04273c76e7
1 changed files with 19 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
content/talk/20220222_siampp/index.md Normal file
View File

@@ -0,0 +1,19 @@
---
title: Automatic Discovery of Implementation Rules for Fast GPU + MPI Operations
date: "2022-02-22T00:00:00Z"
tags: [mpi]
---
*Feb 25th, 5:00 PM MS61 Experiences in Developing GPU Support for DOE Math Libraries*
Invited talk to SIAM Parallel Processing Mini-symposium
### Abstract
Developing a high-performance implementation of a distributed computational kernel for high-performacing computing is increasingly challenging. Systems are composed of heterogenous computational resources, and limited communication performance demands an asynchronous application design. Even if high-performance computation and communication libraries are available. the challenge becomes the best coordination of the provided operations to create an optimal result. This work presents a system that automatically generates design rules for a high-performance implementation of a compound operation provided as a dependence graph. The system searches among valid schedules to determine the fastest arrangement of operations. A post-processing step on the results of the search yields interpretable design rules. The fast implementation can be used directly, or experts can use the design rules to create a high-performance implementation.
### Link
* [slides](/pdf/20220225_siampp.pdf)