Session

DescriptionInter-node dynamic load balancing speeds up scientific simulations on a supercomputer. Given that nodes of emerging supercomputers have a much larger number and variety of processors than previous supercomputers, as well as the continued need for scalability of applications, intra-node load balancing is increasingly needed. Repurposing inter-node load balancing techniques for intra-node load balancing is not practical due to heavyweight processor virtualization and because of the costs incurred from load balancing itself, e.g., data movement from task migration, are more intricate and harder to predict. Node-level load balancing requires novel low-cost load balancing strategies and empirical auto-tuning of such strategies. This mini-symposium brings forward state-of-the-art research involving such node-level load balancing from the level of (1) numerical algorithms via the SLATE library, (2) adaptive and intelligent runtime systems via Charm++’s node-level load balancing, (3) compilers via affinity and loop transformations in LLVM’s OpenMP, and (4) low-level runtime libraries via discussion of the Argobots runtime system. By attending this mini-symposium, one can expect lively discussions, research exchanges, and ideas for future work between the audience and presenters on support for node-level load balancing and its synergy with other techniques within HPC software.