Hunting Down Load Imbalance: A Moving Target.
Christoph Pospiech · ISC · 2015
Load imbalance is known to be a major bottleneck to scalability, particularly when aiming for large parallel partitions. Apart from re-balancing the distribution of the input data, several cures have been proposed. Pretty much all of them assume that the load imbalance is coming from a fixed source. This paper presents an investigation for the climate modelling version CCLM of the weather forecast code COSMO [4]. An adapted MPI trace library is used to collect information about the load imbalance thus introducing a load imbalance measure. Using the visualization software OpenDX [6], this information is correlated to geography and weather forecast results. The resulting pictures show that the locations of high computational load move in space and time. They appear to be correlated to some weather phenomena. In principle this correlation is known for many years [8], but now it has been made visible.