TAU Performance Engineering of HPC and AI Applications Programmed in Julia

The increased complexity of exascale high-performance computing systems compared to previous generations creates difficulties in productively creating performant and portable scientific codes. At the same time, there is growing demand for techniques for memory-safe programming to increase confidence in correctness. One approach to solving these problems is the use of high-level programming languages designed specifically to target scientific computing. One such language is Julia, which has recently been gaining popularity in high-performance computing for government, academic and industrial uses. High-level Julia code is lowered to an LLVM intermediate language and then to native code, and may also incorporate native library code and be part of multi-language workflows. To get the best performance from Julia code requires performance tools capable of spanning these levels of abstraction. To address this problem, this project proposes to extend the TAU Performance System, a widely-adopted suite of HPC performance tools which currently targets primarily C, C++, Fortran and Python to additionally target Julia. Phase I activities will include evaluating and developing proof-of-concept implementations of Julia profiling support for CPU codes and GPU codes using an NVIDIA CUDA backend. Phase II will extend this to multiple GPU platforms and distributed computing models. The target market includes all organizations which develop scientific codes in Julia.