Graphical representations of program performance on hypercube message-passing multiprocessors
Alva L. Couch · 1988
To our senses most commercially available parallel message-passing computers are black boxes of processors, with little external indication of the function of the complex hardware within. The programmer of such a machine must visualize proper program execution in the network while receiving diagnostics only from individual processors. This disguises causal relationships between design and performance. Using the Intel iPSC and NCUBE hypercubes as examples, we show how global performance of a message-passing multiprocessor can be reconstructed from event traces of message passing activity for each constituent processor. We combine event traces from all processors into a single global event trace, and use this to compute a state trace for the entire network of processors. Seecube is an interactive graphical analysis tool for viewing these state traces. States are viewed as collections of scalar execution parameters. The user defines a mapping from parameter values to colors for six different parameter types. Parts of the state of the network are displayed as colorations of otherwise static displays, where parameters are represented as immobile display objects and the object's color codes the value of its parameter. Several different views are provided. Any collection of views can be displayed while a sequencer plays back the state trace on every view simultaneously, in real time proportional to observed event timing. Observing channel state in large hypercube multiprocessors requires an efficient method for representing large-dimension hypercubes in the plane. A hypercube whose dimension is a power of two may be represented in the plane as a set of labeled two-dimensional toroidal grids, where grid nodes are labeled with the node addresses of nodes in the hypercube, and each point in a grid is topologically identified with a unique point in each of the other grids with the same label. As a corollary we give a procedure for constructing a complete set of edge-disjoint Hamiltonian paths and a complete set of embedded edge-disjoint two-dimensional toroidal grids on hypercubes of this size.