This paper presents an overview of an architecture of a multi-micro-processor architecture in which all processors are equivalent. The structure is characterized by a central memory and several caches, and is designed to avoid incoherent data. A fast mechanism of “subcontracting” between one processor and another is described. The execution of two simple programs under this architecture is studied. The parallel algorithms are described and compared with the original (sequential) algorithms; it is shown that the overhead remains small, few extra memories are necessary and synchronization do not slow down the execution unduly. Furthermore, the locality of the parallel algorithm is compared to that of the sequential algorithm and is found to be less good, but in a reasonable way. As well, an estimation of the cost of the coherence keeping mechanism is given using the difference between the miss ratio of the classical cache and the “coherence keeping cache”, respectively.
Unfortunately, ACM prohibits us from displaying non-influential references for this paper.
To see the full reference list, please visit http://dl.acm.org/citation.cfm?id=810653.