TuneDinia
To study the effect of different code optimizations and other tunable parameters on GPU performance it is crucial to have a representative set of tunable applications. Existing benchmark suites, e.g. PARSEC and Rodinia, are straightforward implementations that are not tunable, with hard coded parameters such as the number of threads.
Therefore, many auto-tuning studies only focus tuning on a single application. Studies that do consider multiple applications only tune clock frequencies, or thread block dimensions, because these can be parameterized without significant code modifications. We will create a highly-tunable, representative benchmark suite to be a catalyst for auto-tuning research in general, in addition to being a much more relevant benchmark suite for evaluating the performance of new hardware architectures.
In TuneDinia, we optimized and inserted tunable parameters on the test cases of the Rodinia Benchmark suite. For the purpose of auto-tuning the Kernel Tuner[1] was used. In the first iteration of this project we selected the test cases present on Rodinia's original publication ( https://www.cs.virginia.edu/~skadron/Papers/rodinia_iiswc09.pdf ). In the next steps of this project more test cases will be included and more tunable parameters will be inserted on each of them.
The existing version includes the following test cases :
Backprop
Hotspot
BFS
Leukocyte Tracking
Streamcluster
Kmeans
Needleman Wunsch
Gaussian
SRAD
Tuning
In order to tune the kernels do the following :
- cd tuning
- cd scripts
- ./run_tuner.sh
The run_tuner.sh script will create for each test case :
1) A header file with device targets for compiling a kernel with different parameters on different devices (see https://benvanwerkhoven.github.io/kernel_tuner/user-api.html#kernel_tuner.create_device_targets)
2) A .json file that stores the top (3% by default) best kernel configurations (see https://benvanwerkhoven.github.io/kernel_tuner/user-api.html#kernel_tuner.store_results)
End Goal
Several bugs may exist and not all Rodinia's test cases are tuned up to now. Our end goal is to optimize and tune each test case, so that the end user can use TuneDinia in order to :
- Evaluate the performance of new hardware architectures
- Assess the effect of different code optimizations and other tunable parameters on GPU energy efficiency and performance for a variety of applications belonging to different domains such as Data Mining, Graph Algorithms, Medical Imaging and others.
Documentation
Useful link : https://benvanwerkhoven.github.io/kernel_tuner/user-api.html
Citation
[1]
article = kerneltuner
author = Ben van Werkhoven
title = Kernel Tuner: A search-optimizing GPU code auto-tuner
journal = Future Generation Computer Systems
year = 2019
volume = 90
pages = 347-358
url = https://www.sciencedirect.com/science/article/pii/S0167739X18313359
doi = https://doi.org/10.1016/j.future.2018.08.004
[2]
article = willemsen2021bayesian
author = Willemsen, Floris-Jan and Van Nieuwpoort, Rob and Van Werkhoven, Ben
title = Bayesian Optimization for auto-tuning GPU kernels
journal = International Workshop on Performance Modeling, Benchmarking and Simulation
of High Performance Computer Systems (PMBS) at Supercomputing (SC21)
year = 2021
url = https://arxiv.org/abs/2111.14991
}