Managing Control Asynchrony on SIMD Machines—a Survey

Abstract

Abstract The Single Instruction Multiple Data (SIMD) paradigm has several desirable characteristics from the perspective of massively-parallel algorithms. However, its restricted control organization makes it only useful for a small set of applications that fit this restricted model. The alternative for other applications has been to use Multiple Instruction Multiple Data (MIMD) organizations, which are much more expensive and difficult to use. In addition, MIMD machines are inefficient on fine-grained applications, which require frequent interaction among the processing elements (PEs). This paper surveys research efforts that seek to reduce the inefficiencies associated with managing control-parallelism on SIMD machines. 1 These efforts are organized into three major categories: (1) advances made within the SIMD model; in particular, autonomous features allowing the PEs to locally modify the operation of some instructions, (2) architectures that have attempted to retain the advantages of the SIMD model, while enriching its control organization to allow better support of control-parallelism, and (3) the concurrent-interpretation model: a centralized-control model that shares the control unit at the instruction level (instead of the thread level, as per the SIMD model), allowing efficient support of control-parallelism on a centralized-control organization.