Overcoming the startup time problem in distributed memory architectures

Massively parallel systems are distributed memory architectures consisting of a very large number of autonomously operating, interconnected nodes. The nodes cooperate through message passing via a high-speed interconnection network. The critical issue in such systems is communication latency. A major component of communication latency is the message startup time, i.e. the time it takes to execute the appropriate operating system kernel function for message passing. One of the necessary system optimizations therefore concerns the startup time minimization. The author's approach to this problem is to provide different versions of the kernel for the different modes of operation of the system. All versions have the same interface to the other modules of the distributed operating system and, thus, are interchangeable. This is accomplished by representing the message-passing kernel as an abstract datatype.<>