Publications

A low impact flow control implementation for offload communication interfaces

Barrett, Brian W.; Brightwell, Ronald B.; Underwood, Keith D.

Message passing paradigms provide for many to one messaging patterns that result in receive side resource exhaustion. Traditionally, MPI implementations layered over the Portals network programming interface provided a large default unexpected receive buffer space, the user was expected to configure the buffer size to the application demand, and the application was aborted when the buffer space was overrun. The Portals 4 design provides a set of primitives for implementing scalable resource exhaustion recovery without negatively impacting normal operation. A resource exhaustion recovery protocol for MPI implementations is presented, as well as performance results for an Open MPI implementation of the protocol. © 2012 Springer-Verlag.