Memory conflict resolution in vector supercomputers

This paper presents a simulation study of several conflict-handling methods in shared-memory vector processors. These methods differ with respect to (1) the way memory sections are accessed, (2) the types of conflicts that may occur, (3) whether a conflict halts the vector stream or not, and (4) buffering of pending requests. In the first method, access privilege to a particular section of memory is exclusively granted to each vector stream in its own time phase. This static control strategy prevents contention in the memory interconnect at the expense of a limited capability to adapt to dynamic situations. Two dynamic methods are also considered, in which simultaneous bank conflicts and line conflicts may occur and are resolved at run time. A fourth method helps to identify the relative contribution of the different types of memory conflicts. We also study linked conflicts that occur due to buffering in the time-phased strategy.