الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 182 |  |  |
| | | from | 182 | | |
Abstract
Technology of High Performance Computing (HPC) has improved quickly in latest years. HPC becomes necessary for the efficient processing of the scientific and industrial applications which needs to run within reasonable time. At the same time HPC infrastructure (Computing nodes, networks and storage) is expensive and needs frequently special maintenance. Today, advantages of Cloud-computing such as reliability, scalability and resource-pooling have attracted-researchers to run applications of HPC on the Cloud. To take benefits of HPC and Cloud-computing, it has become reasonable and practical to considerably enhance the running of single application speed-by parallelization of its hotspots calculations with preserving the same level of computations accuracy.
One of the applications which needs applying parallel algorithm in the field of reservoir engineering is 3D (three dimensions) oil reservoir data visualization tool which can render and visualize the output of reservoir simulator (such as pressure, oil saturation and water saturation) in 3D environment to assist the decision maker in statistical analysis, historical matching and recovery of hydrocarbons of the oil reservoir.
In this thesis, we introduce new parallel-techniques of intensive calculations for 3D oil reservoir data visualization tool using Message Passing Interface (MPI), Multi-threaded and hybrid (MPI/Multi-threaded) parallel programming model. The presented
parallelization algorithms can be abstracted into a generalized paradigm for other applications that need parallel techniques for distributed and shared memory models.
Our parallel-techniques for 3D oil reservoir data visualization tool is tested on two platforms, traditional HPC platform and Virtual Cluster on Cloud to enable users of the tool to deploy the suitable parallel technique on their available infrastructure.
Our results indicate that, MPI (Distributed Memory) parallelism is more appropriate with coarse grain data-decomposition technique due to avoid small and frequent-data exchanges among-MPI tasks over the network. Moreover, distributed memory model provides efficient hardware scaling with the Data-grid-scaling. On-the-other-hand, Multi-threading(Shared Memory) parallelism gives high performance with fine grain data-decomposition technique although hardware scalability of shared-memory model is limited due-to its resources-sharing of hardware. Hybrid (shared/distributed memory) approach that uses coarse grain and fine grain data-decomposition techniques concurrently is the best-customed parallelism which is suitable for the nature of 3D grid data-set in addition to use-of mixture system of parallel-processing units(Threads and processors). Our hybrid parallelism on Electronics Research Institute (ERI)-HPC (144 Intel-Xeon cores support Hyper-Threading-Technology) provides on average 284X speedup-over serial-implementation on single P-P-U (Power Processor Unit) of IBM cell B-E (Broadband Engine) and can scale-exceptionally very-well with massive data-sets.