- author = "Jhe-Yu Liou and Muaaz Awan and Steven Hofmeyr and Stephanie Forrest and Carole-Jean Wu",
- title = "Understanding the Power of Evolutionary Computation for {GPU} Code Optimization",
- booktitle = "2022 IEEE International Symposium on Workload Characterization (IISWC)",
- year = "2022",
- pages = "185--198",
- address = "Austin, TX, USA",
- month = "6-8 " # nov,
- keywords = "genetic algorithms, genetic programming, genetic improvement, grammatical evolution, automated program optimization, CUDA, GPU, GEVO, Clang, C++, LLVM IR, SBSE, codes, runtime, scientific computing, computational modeling, search methods, graphics processing units, ADEPT-V1 Smith-Waterman, nvidie P100, nvcc, PTX, computer architecture, epistasisi",
- isbn13 = "978-1-6654-8798-6",
-
DOI = "
doi:10.1109/IISWC55918.2022.00025",
- abstract = "Achieving high performance for GPU codes requires developers to have significant knowledge in parallel programming and GPU architectures, and in-depth understanding of the application. This combination makes it challenging to find performance optimizations for GPU-based applications, especially in scientific computing. This paper shows that significant speedups can be achieved on two quite different scientific workloads using the tool, GEVO, to improve performance over human-optimized GPU code. GEVO uses evolutionary computation to find code edits that improve the runtime of a multiple sequence alignment kernel and a SARS-CoV-2 simulation by 28.9 percent and 29 percent respectively. Further, when GEVO begins with an early, unoptimized version of the sequence alignment program, it finds an impressive 30 times speedup performance improvement similar to that of the hand-tuned version. This work presents an in-depth analysis of the discovered optimizations, revealing that the primary sources of improvement vary across applications; that most of the optimizations generalize across GPU architectures; and that several of the most important optimizations involve significant code interdependencies. The results showcase the potential of automated program optimization tools to help reduce the optimization burden for scientific computing developers and enhance performance portability for domain-specific accelerators.",
-
notes = "See \cite{Liou:2022:arxiv}. Also known as
\cite{9975391}
GEVO mutate LLVM-IR using copy, delete, move, replace, or swap then compile LLVM-IR to PTX and run on GPU by fitness function. pop=256 (elite 4) 80% Crossover 30% mutation. Run for 2 or 7 days.
'modified GEVO’s mutation operator to encode the source code location information'
'ADEPT-V1 [hand optimised] executes approximately 20-30 times faster than ADEPT-V0'.
3 kernels approx 1700 lines of LLVM-IR
Slides: http://www.iiswc.org/iiswc2022/IISWC2022_46.pdf",
