high performance computing on graphics processing units: hgpu.org

Heterogeneous systems are widespread. When neatly used, they enable an impressive performance increase. However, they typically demand developers to combine multiple programming models, languages and tools into very complex programs that are hard to design and debug. Writing correct heterogeneous programs is difficult, achieving good performance is even harder. To help developers, many high-level solutions have been developed, in particular for GPGPU programming (combining a CPU host with one or many GPUs). Most of the time, these solutions generate part of the code that will actually be executed and insulate the programmer from low-level architectural details. They help build efficient and safer programs, but they leave programmers clueless when facing complex bugs or when trying to optimize their application further. In this paper, we propose to enable profiling of heterogeneous applications built using such tools. We present our solution for the SPOC GPGPU framework for the OCaml programming language. Using this framework, we show how it is possible to provide relevant information to the developer focusing on the different parts of GPGPU programs: firstly, by instrumenting the high-level runtime host library of the framework, second, by benefiting from the dynamic generation of co-processor subprograms (GPGPU kernels) in this kind of frameworks to automatically instrument and measure computations, in order to provide developers understandable feedback on their programs. The solution we present, using simple examples, enables profiling of high-level systems in very heterogeneous systems that can combine a CPU host with numerous very different co-processors.