Your timezone:
Wednesday, 14 May 15:30 - 15:55
(Europe/Lisbon)
Session room 3 | Opala III
“Thread Affinity” controls which cores are favoured when distributing threads and processes. Getting the thread affinity right is key for getting the most out an MPI/OpenMP job, and bad affinity can often silently degrade the performance. For existing systems, this is typically configured globally and tweaked/optimized over the life-time of the system. However, when porting to a new system (such as from x86_64 to aarch64 machines), this is no-longer the case. In the early stages of adoption, it can be particularly difficult to detect performance issues caused by bad thread affinity. This technical presentation explores how bad thread affinity can affect the performance of your jobs and will showcase, through a worked example, how Linaro Forge's Thread Affinity Advisor can be used to improve affinity and squeeze the most of your new HPC system.
