Anmerkungen:
Diese Datenquelle enthält auch Bestandsnachweise, die nicht zu einem Volltext führen.
Beschreibung:
Scientific simulations are rarely attempted in a cloud due to the substantial performance costs of virtualization. Considerable communication overheads, intolerable latencies, and inefficient hardware emulation are the main reasons why this emerging technology has not been fully exploited. On the other hand, the progress of computing infrastructure nowadays is strongly dependent on perspective storage medium development, where efficient micromagnetic simulations play a vital role in future memory design. This thesis addresses both these topics by merging micromagnetic simulations with the latest OpenStack cloud implementation while providing a time and costeffective alternative to expensive computing centers. However, many challenges have to be addressed before a high-performance cloud platform emerges as a solution for problems in micromagnetic research communities. First, the best solver candidate has to be selected and further improved, particularly in the parallelization and process communication domain. Second, a 3-level cloud communication hierarchy needs to be recognized and each segment adequately addressed. The required steps include breaking the VMisolation for the host’s shared memory activation, cloud network-stack tuning, optimization, and efficient communication hardware integration. The project work concludes with practical measurements and confirmation of successfully implemented simulation into an open-source cloud environment. It is achieved that the renewed Magpar solver runs for the first time in the OpenStack cloud by using ivshmem for shared memory communication. Also, extensive measurements proved the effectiveness of our solutions, yielding from sixty percent to over ten times better results than those achieved in the standard cloud. ; Aufgrund der erheblichen Leistungskosten der Virtualisierung werden wissenschaftliche Simulationen in einer Cloud selten versucht. Beträchtlicher Kommunikationsaufwand, erhebliche Latenzen und ineffiziente Hardwareemulation sind die Hauptgründe, warum diese ...