A Locality-conscious load-balancer based on negotiations in dynamic unstructured mesh computations

Abstract

Recently hybrid/multi-level parallel programming models are gaining lots of momentum basically because they have proven to provide better scalability, speedup and utilization than any single parallel programming model alone. In such models, load balancing should not only mean balancing the computational loads (as it has always been perceived), but should also mean balancing I/O imbalance as well as synchronization imbalance. In this paper, we propose a broader generic application/language/model independent multi-agent framework for dynamic load balancing. It takes most of the load-balancing burden away from programmers. It is not a library but a runtime support system that is not hardwired to the parallel applications. The framework is intended to handle varying levels of load changes in computations, I/O and/or synchronization throughout the application run and it is an open-architecture that currently supports four multi-level parallel programming models. It has a clean interface to the application, runs in parallel and provides additional functionality such as determination of when to balance load and provide interface to end users. The proposed open-architecture multi-agent load-balancing capability currently makes use of a leading geometric partitioner engine (Chaco) at runtime. A mesh solver may initially create hundreds of lightweight threads, each handling a small submesh by calling Chaco partitioning engine in a pre-processing stage. This partitioner engine might be called again by these light-weight threads if a divide-and-conquer process is deemed necessary when the sub-domain (submesh) served by this thread grows out beyond certain threshold limits and thus creates an imbalance. In the proposed framework, the multi-agent is a set of SMP-based load balancers (agents) that do not have to share any data structure with the parallel application threads. They just monitor and collect system and application data frequently from the outside of the multi-threaded parallel application solver and send adjustments and negotiation plans to the SMP-load balancers and the application threads whenever a need for load balancing arises. The proposed framework has been deployed in four hybrid/multi-level parallel programming models and its capabilities of issuing corrective actions against emerging imbalances were tested in the context of an adaptive mesh refinement application. Experimental results show that the framework is effective in monitoring, tuning and rebalancing emerging computational, I/O and synchronization sources of load imbalance.