Ruprecht-Karls-Universität Heidelberg

Concept and Prototype for a Global Collective Engine to Offload Nonblocking Collective Computation Operations in MPI

Master Thesis by Manuel Abele


With increasing size and complexity of modern high-performance computing systems, their performance demands on inter-process communication grow as well. The means of choice, therefore, is the Message Passing Interface (MPI). Particularly collective communication, where a large quantity of processes can be involved and data is not only transported, but also processed, pose a bottleneck.

Within the scope of this work, a concept to completely offload collective MPI operations of this type is developed. This so-called Global Collective Engine (GCE) is integrated in the EXTOLL interconnect and capable of computing both blocking and non-blocking collective MPI operations with low latency. Beyond the concept, a Verilog prototype is developed and implemented on a Field Programmable Gate Array (FPGA).


« back

back to top