Sender coordination in the Distributed Virtual Communication Machine

The Distributed Virtual Communication Machine (DVCM) is an extensible communication architecture for tightly-coupled clusters of workstations (COWs) connected by high-speed networks. The DVCM is designed for off-the-shelf network interface cards equipped with communication coprocessors. Its main component is an active backplane implemented in firmware running on the coprocessors. This backplane can be extended with modules that implement application-specific-functionality and have access to some of the application´s state. Consequently non-trivial collective computations can be implemented as DVCM extensions. We present a DVCM extension module that provides application-specific network flow control by coordinating the resource-competing components of a parallel application running on an ATM LAN. Our experiments show that this extension module helps eliminate message loss and achieve high link bandwidth utilization when there is significant link contention.