(Description of Seminar 6)
Oct. 6, 1997
On last wednesday, Dr. Seyfarth gave us a lecture about Implementing Glenda on the iPSC/860. Dr. Seyfarth is a professor of Computer Science Department of USM (University of Southern Mississippi.
This software is being developed to provide the capabilities of Linda (Linda is a registered trademark of Scientific Computing Associates), a group of functions capable of utilizing multiple processors which allows the user to more easily program in parallel. Named Glenda, its model was setup to closely parallel that of Linda - with a few exceptions. Glenda is built utilizing Parallel Virtual Machine (PVM 3.x) software system to provide the underlying communications. It supports most of the Linda operations with a few added capabilities to utilize PVM more fully. Glenda 0.91 was funded in part by the National Aeronautics and Space Administration and the USM Center for Ocean and Atmospheric Modeling.
Dr. Seyfarth presents the Model, Architecture, Optimizations, Performance of Glenda system. Tuple is an ordered collection of data, it is not directed to a specific process. Tuples are retrieved based on matching content. The preprocessor converts Glenda calls to C function calls. Reads a hash file to learn how to hash tuples distributing the tuple space. It converts tuple names into integers, converts each tuple component to several function parameters to convey the type.
Hypercube is with up to 128 Intel iPSC860s. Hypercube channels are ezch 2.8 MBps. System resource manager (386) connected at 2.5 MBps. Remote hosts are connected via Ethernet 200 KBps.
iPS support only 1 process per node. Tuple server is actually an interrupt handler in the application. Tuples are hashed to "readezvous" nodes to distribute the tuple service load. Tuple parameter count added to function calls. Hash file is issued to mark tuple. Each nodes has a pool of fixed size buffer. A tuple is stored as a linked collection of buffers. Data format conversion is only applied on the remote host.