We investigate the design of parallel B&B in large scale heterogeneous compute environments where processing units can be composed of a mixture of multiple shared memory cores, multiple distributed CPUs and multiple GPUs devices. We describe two approaches addressing the critical issue of how to map B&B workload with the different levels of parallelism exposed by the target compute platform. We also contribute a throughout large scale experimental study which allows us to derive a comprehensive and fair analysis of the proposed approaches under different system configurations using up to 16 GPUs and up to 512 distributed cores. Our results shed more light on the main challenges one has to face when tackling B&B algorithms while describing efficient techniques to address them. In particular, we are able to obtain linear speed-ups at moderate scales where adaptive load balancing among the heterogeneous compute resources is shown to have a significant impact on performance. At the largest scales, intranode parallelism and hybrid decentralized load balancing is shown to have a crucial importance in order to alleviate locking issues among shared memory threads and to scale the distributed resources while optimizing communication costs and minimizing idle times.
To deal with dynamic load balancing in large scale distributed systems, we propose to organize computing resources following a logical peer-to-peer overlay and to distribute the load according to the so-defined overlay. We use a tree as a logical structure connecting distributed nodes and we balance the load according to the size of induced subtrees. We conduct extensive experiments involving up to 1000 computing cores and provide a throughout analysis of different properties of our generic approach for two different applications, namely, the standard Unbalanced Tree Search and the more challenging parallel Branch-and-Bound algorithm. Substantial improvements are reported in comparison with the classical random work stealing and two finely tuned application specific strategies taken from the literature.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.