Optimal Load Balancing and Assessment of Existing Load Balancing Criteria

TL;DR

This paper introduces a new automatic load balancing criterion and an efficient algorithm to find optimal load balancing points, significantly improving performance in parallel iterative applications.

Contribution

It presents a novel mathematical model for automatic load balancing and a branch-and-bound algorithm to identify optimal load balancing iterations efficiently.

Findings

The proposed criterion outperforms existing automatic criteria in synthetic benchmarks.

Numerical experiments show the new criterion is on average 4.9% faster, up to 17.6%.

State-of-the-art criteria are up to 47.4% slower than the optimal.

Abstract

Parallel iterative applications often suffer from load imbalance, one of the most critical performance degradation factors. Hence, load balancing techniques are used to distribute the workload evenly to maximize performance. A key challenge is to know \textit{when} to use load balancing techniques. In general, this is done through load balancing criteria, which trigger load balancing based on runtime application data and/or user-defined information. In the first part of this paper, we introduce a novel, automatic load balancing criterion derived from a simple mathematical model. In the second part, we propose a branch-and-bound algorithm to find the load balancing iterations that lead to the optimal application performance. This algorithm finds the optimal load balancing scenario in quadratic time while, to the best of our knowledge, this has never been addressed in less than an exponential time. Finally, we compare the performance of the scenarios produced by state-of-the-art load balancing criteria relative to the optimal load balancing scenario in synthetic benchmarks and parallel N-body simulations. In the synthetic benchmarks, we observe that the proposed criterion outperforms the other automatic criteria. In the numerical experiments, we show that our new criterion is, on average, $4.9\%$ faster than state-of-the-art load balancing criteria and can outperform them by up to $17.6\%$. Moreover, we see in the numerical study that the state-of-the-art automatic criteria are at worst $47.4\%$ slower than the optimum and at best $16.5\%$ slower.