TOTAL ECLIPSE: An Efficient Architectural Realization of the Parallel Random Access Machine

In this chapter, we introduce a configurable chip multiprocessor architecture, TOTAL ECLIPSE, for realizing one of the most powerful parallel random access machine (PRAM) variants, the arbitrary multioperation concurrent read concurrent write (MCRCW) PRAM model. In addition to standard arbitrary concurrent read concurrent write (CRCW) PRAM capable of concurrent reads and writes so that in the case of a write arbitrary of the participating threads succeeds, MCRCW provides multioperations that can e.g. sum the values sent by all participating threads into a memory location concurrently. The architecture is optimized for efficient execution of programs containing enough TLP to hide the latency of the intercommunication network and co-exploitation of virtual ILP with TLP but it is also able to execute programs with low TLP efficiently by providing seamless configurability of PRAM threads to non-uniform memory access (NUMA) bunches combining the computational power of two or more threads within a processor core. We will describe the principles of PRAM realization, integration of NUMA bunching to TOTAL ECLIPSE operation, as well as overall architectural structure and operation of the TOTAL ECLIPSE architecture. Performance evaluation by executing simple programs with a clock-accurate simulator is provided and silicon area and power consumption estimations of selected TOTAL ECLIPSE CMP configurations are given. This chapter acts also as a case-driven introduction to novel techniques for parallel architectures, unknown from the theory of sequential architectures.