Datensatz

Zusammenfassung

In this thesis we look at a performance bottleneck of running molecular dynamics
code on GPGPU devices (specifically the CUDA platform), namely random memory
access into global graphics card memory. We offer a solution that involves reordering
memory blocks in order to enable more memory reads from the local cache (shared
memory) instead of the global memory. For determining a memory block order that
sufficiently increases performance a space-filling curve was used.
Significant performance increases for higher filling fractions were observed with a
molecular dynamics simulation that was written from scratch.