Baude, F. (1992) PRAM implementation on fine-grained MIMD multicomputers. Technical Report. Department of Computer Science, Coventry, UK.
Abstract
The more user-friendly tools for expressing data-parallelism are encompassed by the PRAM language mechanisms (mainly, global addressing space, implicit processes synchronization). In this paper, we evaluate the cost for implementing such mechanisms on the only class of parallel architectures that should prove to be really scalable as well as versatile, i.e. MIMD fine-grained multicomputers. Being massively parallel, these architectures are constrained to not exceed a given wire density and also to have fine-grained processing nodes. For these reasons, recent algorithmic and hardware technics for hidding communication latency of PRAM emulations by parallel slackness do not apply. We consider an abstraction of these machines by reasoning in terms of the actor model of concurrent computation. Given the architectural constraints raised by the chosen class of architectures, the idea for implementing any PRAM program is to transform it into an actor program such that it can be efficiently implemented on the target machine via mapping technics such as for exemple graph embedding. This requires the underlying communication graph of the actor program to be predictable and sparse. This transformation is grounded on probabilistic theoretical simulations of PRAMs on parallel machines based on sparse communication networks.
Actions (login required)
![[img]](http://eprints.dcs.warwick.ac.uk/style/images/fileicons/application_pdf.png)
