Optimising Task Placement for High Performance Computing
When a complex computation is split up into work for different processors these processors often have to share data and communicate results as the computation progresses. The current state of the art doesn’t consider these communcation requirements when scheduling these tasks. We developed a technique to automatically analyse how the task shares data between processors and use this to optimise the placement of threads and data, resulting in a 10% speed increase for amenable programs.
Keywords
HPC, Memory Communication, Process Communication
Staff
[Christopher Brown]{cmb21} and [Ryan Kirkpatrick]{rw86}