Open MP Extension for Multithreaded Computing with Dynamic SMP Processor Clusters with Communication on the Fly

This paper presents a possible extension of the Open MP library for programming parallel multithreaded computations in the architecture of dynamic SMP clusters with communication on the fly. Dynamic SMP clusters are composed of processors directly connected to the same local shared memory modules with the composition of the clusters arranged at program runtime. Inter-processor communication in such clusters is based on a new "communication on the fly" paradigm. It enables direct communication between processor data caches and eliminates many data read/write transactions concerning memory modules. New functions are presented that should be introduced to standard Open MP library to enable writing parallel multithreaded programs with communication based on the new architectural features. These functions are illustrated using the example of a program in C for parallel matrix multiplication based on data decomposition into quarters