Artifact Description: Performance Modeling of Streaming Kernels and Sparse Matrix-Vector Multiplication on A64FX
- Christie Alappat, RRZE University of Erlangen-Nürnberg, christie.alappat@fau.de
- Jan Laukemann, RRZE University of Erlangen-Nürnberg, jan.laukemann@fau.de
- Thomas Gruber, RRZE University of Erlangen-Nürnberg, thomas.gruber@fau.de
- Georg Hager, RRZE University of Erlangen-Nürnberg, georg.hager@fau.de
- Gerhard Wellein, RRZE University of Erlangen-Nürnberg, gerhard.wellein@fau.de
- Nils Meyer, University of Regensburg, nils.meyer@ur.de
- Tilo Wettig, University of Regensburg, tilo.wettig@ur.de
The A64FX CPU powers the current #1 supercomputer on the Top500 list. Although it is a traditional cache-based multicore processor, its peak performance and memory bandwidth rival accelerator devices. Generating efficient code for such a new architecture requires a good understanding of its performance features. Using these features, we construct the Execution-Cache-Memory (ECM) performance model for the A64FX processor in the FX700 supercomputer and validate it using streaming loops. We also identify architectural peculiarities and derive optimization hints. Applying the ECM model to sparse matrix-vector multiplication (SpMV), we motivate why the CRS matrix storage format is inappropriate and how the SELL-C-σ format with suitable code optimizations can achieve bandwidth saturation for SpMV.
- Compilation: gcc
- Binary: aarch64
- Hardware: A64FX (FX700)
- Publicly available?: yes
Check out this repository.
git clone https://github.com/RRZE-HPC/pmbs2020-paper-artifact
We ran on a FX700 system (A64FX architecture) at 1.8 GHz with 48 cores per CPU.
The results should be reproducible on any FX700 system.
You can find a detailed state description of the system we used for running the experiments in machine_state.txt.
- GCC 10.1.1
You can find a detailed state description of the system we used for running the experiments in machine_state.txt.
For SpMV we use matrices from SuiteSparse Matrix Collection (https://suitesparse-collection-website.herokuapp.com).
None necessary.
To validate our results use the following commands.
Download script and benchmark codes:
git clone https://github.com/RRZE-HPC/pmbs2020-paper-artifact
cd pmbs20-paper-appendix/
The description for compiling and running can be found in the corresponding README.md files of each benchmark.
Navigate to the in-core_instructions directory:
cd single_core/in-core_instructions/
Run the validation accordingly to the README file and compare the results to Table II in the paper.
We provide code to reproduce the single-core streaming kernel and 2d five-point benchmarks for shown in this publication.
Navigate to either one of the directories:
cd single_core/[stream|stencil]/
Run the validation accordingly to the README file and compare the results to Table III in the paper.
We provide code to reproduce the multi-core streaming kernel and 2d five-point benchmarks shown in this publication.
Navigate to either one of the directories:
cd multi_core/[stream|stencil]/
Run the validation accordingly to the README file and compare the results to Figure 4 in the paper.
To reproduce the results of the SpMV kernels in this work, navigate to the SpMV directory:
cd spmv/
Run the benchmarks according to the README file and compare the results to Figure 5 in the paper.
None
None