Distributed Computation Environment Tools

The objective of this task is to develop application utilities and interface tools for enhancing the ease of application development and the efficiency of resource utilization in a distributed computing environment.

Approach:

The objective of enhancing the distributed computing environment is approached at two levels: the application level and the system level. At the application level, the focus is on development both of parallel algorithms and of utlities for data-parallel-specific applications. At the system level, the focus is on development of mechanisms for inter-process control and resource mapping. The tools employ both the MPI protocol for the distributed-memory architectures and threads for shared-memory architectures, in order to achieve platform independence while preserving optimal performance.

Accomplishments:

• Application Level -- A Global Imageclass Library (C++) employing MPI was completed, and it is available for the Cray T3D and for a heterogeneous workstation environment. An iterative Monte Carlo radiosity solution for a full-multiple-scattering problem was implemented employing both threads and MPI.
• System Level -- The IPC Library, an inter-process communication protocol library, developed for a physically-distributed-memory architecture, was upgraded to employ threads in a shared-memory architecture.

• MPI-based Global Imageclass Library -- This library enables platform-independent parallel implementations for applications with multiple data-parallel functions.
• Distributed Comet coma-illumination modeling -- A paper on "Modeling Cometary Coma with a Three-Dimensional, Anisotropic-Multiple-Scattering Distributed Processing Code" will be presented at the Parco '97 (Parallel Computing Conference) in September, 1997, in Germany. For the PostScript display of the paper, press here.

• Thread-based Inter-Process Communication Library -- This library allows pointers to be passed among processes, reducing communication time significantly.

• The Comet coma-illumination model code was developed for a distributed-memory architecture employing MPI, and for a shared-memory architecture employing threads. The code was benchmarked on four types of computational platforms for scalability and communication efficiency:
  • SGI Onyx (shared memory)
  • SGI Origin (virtual shared memory)
  • Cray T3D (distributed memory)
  • heterogeneous workstations (SGIs and Suns)
• A volume renderer for semi-transparent objects will be implemented for visualizing the phenomena constructed by the Comet coma-illumination model.
• The thread-based inter-process communication library was applied in executing MISR (Multi-angle Imaging Spectro-Radiometer) data processing software in a multi-processor architecture by decomposing the software into multiple processes and dynamically mapping them to the processors. The distributed application will be ported to a 32-node SGI Origin at Langley Center (NASA) and efficient process-to-processor mappings will be studied.