​

Science of Software Development

​

a summary of our position at SSSD-U, 2021

Ganesh Gopalakrishnan, with

members of the X-Stack Project ComPort

https://xstack-fp.github.io/

​

Summary Position

• Basic correctness-checking tools are lacking for accelerators
• Need community effort

Science of Software: Need for Concurrency Analyzers

• Need for tools to check for Concurrency Errors
• Rising use of accelerators
• GPUs power HPC and ML
• GPUs are also central to encryption schemes
• It is easy to introduce data races into high-performance GPU codes
• YET
• There isn't a data race checker for GPUs (source: discussions with academic experts and lab researchers; plus own experience)
• Tools like CUDA-Memcheck do not check for global memory races
• Academic tools in this space do not last (fast-moving support technologies)
• The community has to come together to create basic concurrency correctness checking tools

Science of Software: Need for Numerical Analyzers

• Need for tools to check for Numerical Errors
• Rising use of accelerators
• GPUs power HPC and ML
• GPUs are also central to encryption schemes
• It is easy to introduce exceptional numbers into GPU computations
• NaNs, Overflows, …
• NaNs may well be flowing into HPC and ML codes today
• Insights from the results of FPChecker (Laguna, LLNL)
• YET
• There aren't tools that can look for such basic numerical errors
• Those tools that exist are limited by what CUDA subsets are supported
• Tools for non-Nvidia GPUs can help, but aren't present
• The community has to come together to create basic numerical correctness checking tools

Some Areas to Emphasize

• Formal Methods
• Essential for "Science"
• Testing, including Fuzz Testing
• Relevant coverage
• Testing AI Software; AI Software for Testing
• Two-way street
• Correctness via Performance Portability Layers
• Opportunities for creating abstractions that bridge concurrency models
• Timeliness
• Old SW Correctness Approaches Need Major Redo / Rethink