Optimal load balancing and assessment of existing load balancing criteria

• We propose a new optimal automatic load balancing criterion for parallel iterative applications. introduce an algorithm finding the time at which dynamic should be triggered. Our outperforms state-of-the-art criteria by up to 17.6%, while being, on average, 4.9% faster. show that are most 42% slower than optimum. Parallel applications often suffer from imbalance, one of critical performance degradation factors. Hence, techniques used distribute workload evenly maximize performance. A key challenge is know when use techniques. In general, this done through criteria, trigger based runtime application data and/or user-defined information. first part paper, we novel, derived simple mathematical model. second part, branch-and-bound find iterations lead This finds scenario in polynomial while, best our knowledge, it has never been addressed less exponential time. Finally, compare scenarios produced relative synthetic benchmarks and N-body simulations. benchmarks, observe proposed other criteria. numerical experiments, is, faster can outperform them 17.6%. Moreover, see study worst 26.43% optimum 10% slower.