Data Neighboring in Local Load Balancing Operations

Distributed load balancing strategies are burdened with a high communication latency compared to centralised strategies. This latency is originated due to a necessary status profile reported from each computing node to every node, (all to all operation involving expensive synchronization costs). This handicap is lessened due to a certain locality character imposed by the way these strategies are implemented. Many scientific applications are affected by specific communications patterns, these patterns establishes relationships among neighboring data configurations. Application performance is severely affected when a migration process is carried out without taking into account this data neighborhood relationship. It is our main aim to evaluate the performance gains obtained when load balancing strategies are complemented with locality awareness techniques and applied to such applications. Doing this we ensure that migrations preserve as much as possible the data locality relationship aforementioned. Load balancing and preservation of locality issues were achieved by means of a multithreaded environment called AMPI. This framework makes possible to embed a MPI process into a user level thread. Key-Words: -Dynamic Load Balancing, Local Strategies, Data parallelism, virtual processor.