The International Conference for High Performance Computing, Networking, Storage and Analysis
Load Balancing Scientific Applications.
Student: Olga T. Pearce (Texas A&M University)
Advisor: Nancy M. Amato (Texas A&M University)
Abstract: Optimizing high-performance physical simulations to run on ever-growing
supercomputing systems is challenging. The dynamic behavior of large modern parallel simulation codes can lead to imbalances in computational load among processors. Load imbalance is particularly expensive at scale, because hundreds of thousands of idle processors may wait on a single overloaded processor. Future machines will support even more parallelism and efficiently redistributing and balancing load will be critical for good performance.
In this thesis, I address how to evaluate load imbalance and
make its correction affordable. I developed a model for comparison of load balance algorithms in the context of a specific application imbalance scenario,
enabling the selection of a balancing algorithm that will minimize overall runtime. I provide an accurate and fast method to balance the load in simulations with highly non-uniform density. I devised a framework for decoupling the load balancer from the application, enabling asynchronous load balancing.