high performance computing on graphics processing units: hgpu.org

GPUs (Graphics Processor Units) are being increasingly embraced by the high performance computing and computational communities as an effective way of considerably reducing application execution time by accelerating significant parts of their codes. CUDA (Compute Unified Device Architecture) is a new technology developed by NVIDIA which leverages the parallel compute engine in GPUs. However, the use of GPUs in current HPC clusters presents certain negative side-effects, mainly related with acquisition costs and power consumption. rCUDA (remote CUDA) was recently developed as a software solution to address these concerns. Specifically, it is a middleware that allows transparently sharing a reduced number of CUDA-compatible GPUs among the nodes in a cluster, reducing acquisition costs and power consumption. While the initial prototype versions of rCUDA demonstrated its functionality, they also revealed several concerns related with usability and performance. With respect to usability, the rCUDA framework was limited by its lack of support for the CUDA extensions to the C language. Thus, it was necessary to manually convert the original CUDA source code into C plain code functionally identical but that does not include such extensions. For such purpose, in this document we present a new component of the rCUDA suite that allows an automatic transformation of any CUDA source code into plain C code, so that it can be effectively accommodated within the rCUDA technology.