On memory contention problems in vector multiprocessors

Memory interleaving considerably increases memory bandwidth in vector processor systems. The concurrent operation of the processors can produce memory bank conflicts and hence alter the memory bandwidth. Total or steady state performance for vector operations in a memory system is studied. Many methods of resolving memory bank conflicts are proposed and compared. Analytical results on the resulting effective bandwidth are presented for one of them and the others are described by exhaustive simulations. Some nonintuitive results are obtained on how conflicts depend on the size of the architecture, the number, the stride and the length of the vectors, the register length assigned by each processor to vector components.<>