Innovative and efficient communication methods for system area networks

A new trend is emerging to replace massively parallel machines with clusters built from Commercial Off The Shelf (COTS) components. Clusters typically consist of standard compute nodes and an interconnection network. In order to achieve high efficiency, a parallel application is very dependent on the communication system. Traditional communication interfaces will not let an application exploit the given resources due to overburdened protocols. System Area Networks (SAN) have been developed with the main purpose of supporting user-level communication (ULC) systems. User level communcation bypasses the operating system from the critical communication path. One aspect, however, still needs to be solved: How applications will benefit directly from available resources and in which way protocols can be developed to directly support existing applications. This is one of the reasons why traditional networks using standard protocol stacks remain preferred, thus offering upgrade compatibility. This thesis addresses existing communication principles and provides new protocols for efficient communication. It also provides message passing protocols for the new Atomic Low Latency (ATOLL) System Area Network. This network on a chip solution was analyzed and an extended design was developed which will enhance the current implementation to enable protocol offloading for better resource utilization. Finally, this thesis will present a new middleware layer, which will enable the replacement of traditional networks by providing compatible protocols at binary level. This middleware layer will offer the same semantics at much higher performance. As will be shown, standard existing applications will experience much better performance with an improvement in the order of a magnitude.