Load balancing using semi-private memory

The author introduces the concept of semi-private memory as an architectural means of load balancing. In distributed, message-passing systems, each processing element is typically associated with a private input queue and a private local memory. The semi-private memory concept, however, allows each processing element to service a small, unique subset of the queues and memories that overlaps with the subset serviced by neighboring processing elements. Thus, when a hot-spot occurs, or when a region of higher work occurs, service can be skewed towards that hot-spot or region. This type of architectural load balancing would be useful when the overhead incurred by dynamic software load balancing is prohibitive. Simulations are presented that illustrate the possible performance difference between private and semi-private memory. Higher processor utilization was achieved and hot-spots were cooled-off using semi-private memory