Efficient routing functions for distributed memory multiprocessors

Lines connect the coordinator with its partners. CORDIS does not always give exact coordinates for projects before 2014. These points are placed at city or country level.

Project objective

Massively parallel computers with thousands of processors constitute one of the most viable approaches for the achievement of teraflop performances. These machines usually employ a direct network for the interconnection of their processors. Experimental results show that the communication overhead has been one of the major obstacles for the achievement of pick performances. Last generation multiprocessors employ at each node a special unit, the router, that can handle the communication between distant nodes witout interrupting the processor. Each router realizes a routing function that locally decides how to forward each incoming message to the appropriate output channel. There are several key points that have to be addressed for the choice of a routing function. We propose a thorough investigation of networks that are susceptible to the realization of routing functions that possess several desirable properties, and the derivation of general techniques for the constuction of such routing functions. This project has direct practical interest for the construction of routers and routing functions for distributed memory multiprocessors.

Original text from CORDIS.

Participants

• Université de Paris XI (Université Paris-Sud) · Châtenay-MalabryCoordinatorFrance
• Not availableCity levelGreece

Links

• View on CORDIS