This repository maintains the latest version of this tutorial.
Please cite this tutorial using the following Zenodo DOI information:
Date/Time: September 22nd
- 12:15 to 3:45 PM EDT
- Held virtually via the HPEC video platform
- HPEC Preliminary Agenda
- HPEC provides a complementary registration category, so please join us! Select the “Exploring Graph Analysis for HPC with Near-Memory Accelerators” tutorial when registering, so we know how many attendees will join.
The growth of applications related to machine learning, drug discovery for precision medicine, and knowledge graphs has led to a surge in the usage of graph analysis concepts with high-performance computing platforms. At the same time, recent APIs like GraphBLAS have introduced the concept of using “traditional” linear algebra operators to work with large data sets represented as graphs.
This tutorial provides both an introduction to the concept of near-memory hardware accelerators to support high-performance graph analytics as well as an in-depth hands-on session with one specific near-memory accelerator architecture, the Lucata Pathfinder-S. Near-memory architectures like the Pathfinder are well-suited for graph analytics due to their focus on keeping computation very close to data. This strategy reduces performance penalties commonly associated with traversing graphs or other sparse data structures.
Attendees will learn about possible use cases for graph analytics for HPC applications, and they will be able to use a 32-node Pathfinder system to execute HPC-relevant code examples. GraphBLAS support for the Pathfinder will be discussed, and attendees will have an opportunity to work through several benchmarks built for the Pathfinder using the latest GraphBLAS API.
- Jeffrey Young, Darryl Bailey, Aaron Jezghani, Patrick Lavin, Will Powell - Georgia Tech
- Janice McMahon, Jason Riedy - Lucata Corporation
- Srinivas Eswar - Argonne National Lab
- Attendees should understand HPC-relevant application spaces where graph analysis can be used to process large data sets, including ML applications, knowledge graphs, and emerging fields like drug discovery.
- Attendees will learn the basics of GraphBLAS and how it provides a bridge between traditional BLAS libraries and graph analysis applications.
- Attendees will learn how to program and run simple Cilk-based examples based on SpMV and GEMM that can be run on the Lucata Pathfinder-S system.
- Attendees should gain an appreciation for what a GraphBLAS application looks like, and how this can be used with novel architectures, including the Pathfinder-S system.
| **Time (All Times EDT) | Topic | Presenters | Notes |
|---|---|---|---|
| 12:15 | Tutorial Introduction | Jeff | Slides |
| 12:35 | System Overview of the Pathfinder System and Ecosystem | Janice | Slides |
| 1:00 | Programming Basics, SAXPY | Jeff | Slides, NB #1, NB #2, NB #3 |
| 1:45 | BREAK | ||
| 2:15 | Workflow Discussion | Jeff | Slides, NB #4 |
| 2:45 | GraphBLAS | Jason | Slides, NB #6 |
| 3:25 | Profiling/Timing; Hands-on Investigation | Jeff | Slides, NB #5 |
| 3:45 | Wrap-up |
Please see the Tutorial Instructions for specific details on how to access the notebooks hosted here under the code folder.
This tutorial is geared towards using the 22.09 version of the Lucata programming framework. Please see the programming manual (requires RG and GT account) for more information.
This tutorial information is funded in part by NSF CNS #2016701, “Rogues Gallery: A Community Research Infrastructure for Post-Moore Computing”. We are greatly appreciative of the assistance from Lucata for slides that we have included as well as general support and assistance in supporting the Pathfinder platform.
The initial version of the tutorial material was developed by Janice McMahon with help from Jason Riedy and Jeffrey Young for the notebook design. Please see the AUTHORS file for the updated list of contributors.