Effizienter Nachrichtenaustausch auf speichergekoppelten Rechnerverbundsystemen mit SCI-Verbindungsnetz

A cluster is a group of independent compute nodes which are coupled by an interconnection network. They can be used to execute parallel applications, consisting of processes which communicate via the interconnection network. This work deals with clusters coupled by a high-performance, memory-coupling Scalable Coherent Interface (SCI, IEEE 1596 standard) interconnect. SCI enables transparent access to remote memory locations. This platform serves as a basis for the implementation of a message-passing programming model using the Message Passing Interface (MPI) application programming interface specification. A number of new techniques is developed and applied for this implementation to make a maximal fraction of the potential performance available to the MPI-based parallel application. The upper limit of the communication performance of an SCI interconnect with n-dimensional torus topology is analytically determined for different typical communication scenarios. The required variants of the basic point-to-point communication protocols are designed accordingly. They are evaluated by a new multi-level efficiency model. Next to this, new communication algorithms and protocols for collective operations (which involve an arbitrary number of communicating processes) are presented. They achieve a multiple of the performance of generic techniques, and also surpass the performance of an available SCI-optimized implementation. An innovative technique for the transmission of non-contiguous memory areas increases the transfer bandwidth for this application of MPI derived data types. Up to 100% of the transfer bandwidth of contiguous memory areas have been achieved, while reducing the working set size and thus increasing the locality. For one-sided communication, means for communication and synchronization have been developed. They perform well even for loosely coupled, non-phased communication scenarios by the use of asynchronous synchronisation. Finally, advanced communication protocols and new data transfer mechanisms are shown to make efficient overlapping of communication and computation possible. This allows to hide up to 90% of the communication time. The software package that has been developed with this work merges the new techniques and algorithms into an MPI library of outstanding performance and flexibility. It is freely available including the source code. Therefore, it can easily be used for the efficient employment of clusters and for transferring the findings to other areas of parallel processing