Session

DescriptionScientific and technical computing has been facing challenges posed by hardware trends characterized by increasing hardware parallelism, cost of data movement, specialization and heterogeneity. To fully utilize the potential of upcoming hardware, efficient task parallelization frameworks are required. Successful frameworks need to be capable of efficiently expressing fine-grained concurrency, scheduling microsecond-granularity tasks and data movement, and tightly integrate with communication libraries to allow critical-path optimizations in a fully asynchronous schedule. Most of the current tasking frameworks have either struggled to provide low overhead, or have been mainly targeting accelerators, and integration with communication standards like MPI is still an active research area. Molecular dynamics (MD) has been facing these challenges for a decade due to the typically fixed problem size motivating the strong scaling needs, and very high, sub-millisecond, iteration rates. Hence, the experience in this field can be beneficial to the broader community likely to face similar challenges on future exascale architectures. In this session experts in runtimes and developers of flagship MD codes GROMACS and NAMD will to provide a cross-disciplinary view of state-of-the-art in tasking frameworks, APIs, as well as algorithmic and implementation advances made in tackling fine-grained task scheduling challenges while strong-scaling MD on heterogeneous architectures.