Managing the recursive generation of tasks in a transputer network

Abstract

Computational constraints in image processing and image analysis, take advantage of a parallel implementation performed on a distributed memory parallel architecture; the following one is based on tranputers. This paper proposes a parallel implementation based on the “processors farm” method, of the splitting step of Kd-tree algorithms (split & merge); it is computationally expensive, but involves a large number of independant tasks. The project sets out to implement it on a on-board real time system.

The dynamic comportment of the network is formalized; both master and slave processes are described. A communication strategy is performed which stops the deadlock happening in the return communication network. The recursive generation of sub-tasks may create some starvation points. When new sub-tasks are created, they have to be to idle workers. Two approaches are proposed:

• - firstly, when a starvation point appears, a total starvation is created;

• - secondly, the master does not stop its “farming”, but manages the network temporal comportment and “farm” new sub-tasks to idle workers.

In the end, some experimentals results, performed on several networks composed from 2 till 12 transputers, are presented.