Vivaldi: a Python-like domain-specific language for volume rendering and processing on distributed multi CPU-GPU systems

Abstract

In this thesis, a Python-like domain-specific language alled Vivaldi is proposed. Vivaldi is based on Python, but can also provide parallel volume rendering and processing on distributed heterogeneous systems. Nowadays, visualization requires high performance for processing large data and creating high quality images. Therefore, computing systems have also advanced to meet this requirement; one example is graphics processing unit (GPU) clusters. However, even though high-performance systems exist, effective utilization of these systems is not easy for non-experts such as domain scientists and researchers because they require expert programming skills and a significant amount of software development for parallel programming, distributed systems, and heterogeneous architecture. One aim of Vivaldi is to minimize these required programming skills using a domain-specific language for volume rendering and visualization that is Python-like and platform independent. In this language, a parallel visualization model, virtual shared memory model, and platform independent architecture for distributed heterogeneous systems are proposed. The parallel visualization model provides a simple means to implement visualization pipelines. The virtual shared memory model enables the use of cluster memory without Message Passing Interface (MPI) and CUDA. Finally, the platform independent design integrates central processing unit(CPU)s and GPUs into a common, domain-specific language. Vivaldi code was compared to C++ implementations to evaluate its performance according to number of lines, performance, and scalability. The results show that Vivaldi achieved comparable scalability and performance while requiring much less programming effort.