PackStealLB: A scalable distributed load balancer based on work stealing and workload discretization

The scalability of high-performance, parallel iterative applications is directly affected by how well they use the available computing resources. These are subject to load imbalance due nature and dynamics their computations. It common that high performance systems employ periodic balancing tackle this issue. Dynamic algorithms redistribute application’s workload using heuristics circumvent NP-hard complexity problem However, scheduling must be fast avoid hindering application when distributing on large distributed environments. In work, we present a technique for low overhead, quality decisions applications. relies combined information paired with algorithms. An initial step among agents group tasks in packs similar minimize messages them. This used our algorithm, PackStealLB, its distributed-memory work stealing heuristic. Experimental results showed PackStealLB able improve molecular benchmark up 41%, outperforming other most scenarios over almost one thousand cores.