This repository contains a walkthrough on the High-Performance Linpack (HPL) and High-Performance Conjugate Gradient (HPCG) benchmarks.
Prerequisites
- Access to a compute resource running Linux
- Command-line utilities: screen, git
- Compiler for C/C++:
- [Debian/Ubuntu/Mint] apt-get install build-essential
- [CentOS/RHEL] yum install gcc gcc-c++ make
| Setup | Download files required for the lesson | |
| 00:00 | 1. Introduction |
Why are HPL and HPCG used as benchmarks?
What is a linear solve used for? |
| 00:10 | 2. Computational Kernels | What are memory and compute-bound kernels? |
| 00:50 | 3. BLAS |
What blas should I link with?
How can I get all these libraries installed and working together? What are the limitations of blas and lapack? |
| 01:20 | 4. HPL |
How do dense matrix algorithms divide work using MPI?
What are the key factors affecting their performance? Why is peak performance a reasonable target for HPL? |
| 01:50 | 5. HPCG |
How do sparse matrix algorithms divide work using MPI?
What are the key factors affecting their performance? Why is peak network bandwidth a reasonable target for HPCG? |
| 02:08 | 6. Summary | How can I demonstrate my awesome cluster configuration? |
| 02:22 | 7. Using GPUs | What modifications to the benchmarks are required to utilize GPUs? |
| 02:49 | 8. Using Infiniband | How does infiniband hardware differ from ethernet? |
| 02:59 | 9. Scripting OpenStack | |
| 03:24 | Finish |
The actual schedule may vary slightly depending on the topics and exercises chosen by the instructor.