Message-Passing on PCI-SCI Network Interfaces

SCI has many advantages compared with other modern high speed interconnections that makes it very attractive for building costeffective clusters from standard computer systems and PCI-SCI bridges. In this paper we present some ideas for new architecture of such communication hardware, which will combine the benefits of SCI distributed shared memory with high effectiveness and low latencies of protected user level DMA of the Virtual Interface Architecture (VIA). The final goal of our research is to develop PCI-SCI bridge with completely new design which will be very suitable for building inexpensive clusters optimized for message passing applications. Scaleable Coherent Interface (SCI)[1] is modern communication technology with very high bandwidth, extremely low latency, scalable architecture and support of distributed multiprocessing that allows building of large systems out of many inexpensive units. SCI acts like a modern equivalent of processor/memory/IO bus and Local Area network, which uses a single address space to specify data, as well as its source and destination when being transported and in this way implements distributed shared memory with cache coherency. Thus, when two tasks share data using SCI, the data remains stored in ordinary variables, with ordinary memory addresses, at all times and processor instructions like Load and Store suffice to address the data for doing computation with it. Load and Store instructions are highly optimized in all