guest ::
| Home > Publications database > Optimization of Output Bandwidth from an Intel Paragon |
| Home > Publications database > Optimization of Output Bandwidth from an Intel Paragon |
| Report | FZJ-2015-03204 |
;
1994
Concurrent Supercomputing Consortium
Pasadena, CA
Report No.: CCSF-44
Abstract: We examine the question of how to get maximum output bandwidth from a "real" application running on many nodes of a Paragon. We consider synchronization, flow control, message passing and partition management, and the use of the PARtools performance monitoring and visualization environment to understand and thus optimize these aspects.The application is a "digital VCR": each processor has in memory several compressed video frames, and these are to be decompressed, sent to the Hippi device, and output to a framebuffer at the fastest possible rate. While the Hippi processor is capable of 80 Mbyte/sec, the aggregate bandwidth from the computational processors is 21 Mbyte/sec.In this communication-bound application, we find it optimal to have the communication device (the Hippi processor) to have control, with other processors acting as servers; rather than the processors having distributed control and the Hippi processor acting as server.
|
The record appears in these collections: |