SALT: An LLVM-based Source Analysis Toolkit for HPC

Developers of applications for high performance computing systems need tools to guide code development and optimization so that they can make efficient use of limited computational resources. The upcoming exascale generation of supercomputers feature a variety of processor and accelerator architectures which have driven the adoption of complex libraries for performance portability which make use of recently released language features of C++ and Fortran. Applying performance tools to these libraries and applications requires a high level of expertise which serves as a barrier to adoption. The TAU Performance System is a suite of performance tools which could be used for such applications, but is a complex, expert-level tool. To address this problem, we will harden and improve the usability of TAU by building a next-generation source analysis toolkit (called SALT) based on LLVM compiler technology, a compiler widely adopted both inside and outside the high-performance computing field. The SALT parsers will appear to build systems as if they were the underlying LLVM compilers and their use will simply require changing the environment variable specifying the compiler before building the application. No customization or workarounds for differences between the analysis parser and the compiler will be needed. In this way, performance tools developed with SALT will be made accessible to non-expert users. During Phase I, the new SALT framework will be developed and integrated with TAU. Proof-of-concept implementations will be developed and evaluated for several components. A new, efficient database format will be designed, and a C++ parser supporting all modern versions of the language will be developed using the new format. A simple instrumentation API will be developed, and a code instrumentor and library wrapper generator will be developed using the API. Easy-to-use command line and graphical interfaces will be developed for these components, and they will be tested and evaluated with libraries and applications being developed for exascale systems. The Council on Competitiveness reports that over two-thirds of U.S. industry representatives claim their HPC applications could utilize a 10x increase in computing capability, and over one-third could use a 1000x increase. The affordable performance engineering products developed through this SBIR project will fill a crucial need for improved compute capability utilization by improving software scalability and developer productivity, ultimately accelerating the pace of research and development.