Towards Informed Partitioning for Load Balancing: a Proof-of-Concept

TL;DR

This paper introduces informed partitioning, a new method for load balancing in particle simulations that uses computation evolution to improve performance and reduce balancing calls, demonstrated through a proof-of-concept study.

Contribution

It proposes a novel geometric partitioning technique guided by particle velocity and an effort metric for dynamic load balancing, with a proof-of-concept evaluation on N-Body simulations.

Findings

Increased performance in two out of three simulations by up to 76%.

Achieved marginal slowdown of only 3% in one experiment.

Effort metric effectively ranks partitioning techniques during simulations.

Abstract

Most parallel applications suffer from load imbalance, a crucial performance degradation factor. In particle simulations, this is mainly due to the migration of particles between processing elements, which eventually gather unevenly and create workload imbalance. Dynamic load balancing is used at various iterations to mitigate load imbalance, employing a partitioning method to divide the computational space evenly while minimizing communications. In this paper, we propose a novel partitioning methodology called ``informed partitioning''. It uses information based on the evolution of the computation to reduce the load balancing growth and the number of load balancing calls. We illustrate informed partitioning by proposing a new geometric partitioning technique for particles simulations. This technique is derived from the well-known recursive coordinate bisection and employs the velocity of the particles to guide the bisection axis. To properly compare the performance of our new method with existing partitioning techniques during application execution, we introduce an effort metric based on a theoretical model of load balanced parallel application time. We propose a proof-of-concept of informed partitioning, through a numerical study, on three N-Body simulations with various particle dynamics, and we discuss its performance against popular geometric partitioning techniques. Moreover, we show that our effort metric can be used to rank partitioning techniques by their efficiency at any time point during the simulation. Eventually, this could be used to choose the best partitioning on the fly. In the numerical study, we report that our novel concept increases the performance of two experiments out of three by up to 76% and 15%, while being marginally slower by only $3\%$ in one experiment. Also, we discuss the limitations of our implementation of informed partitioning and our effort metric.