high performance computing on graphics processing units: hgpu.org

We present the Hadoop Online Ray Tracer (HORT), a scalable ray tracing framework for general, pay-as-you-go, cloud computing services. Using MapReduce, HORT partitions the computational workload and scene data differently than other distributed memory ray tracing frameworks. We show that this unique partitioning significantly bounds the data replication costs and inter-process communication. Consequently HORT is fault-tolerant and cost-effective when rendering large-scale scenes (i.e., scenes that do not fit into local memory) without specific or dedicated high performance infrastructure. Our experiments demonstrate this scalability and fault tolerance using several CPU and GPU instances on Amazon AWS with the Hadoop open-source implementation of MapReduce.