Kiloprocessor Extensions to SCI

To expand the Scalable Coherent Interface´s (SCI) capabilities so it can be used to efficiently handle sharing in systems of hundreds or even thousands of processors, the SCI working group is developing the Kiloprocessor Extensions to SCI. We describe the proposed GLOW and STEM kiloprocessor extensions to SCI. These two sets of extensions provide SCI with scalable reads and scalable writes to widely-shared data. This kind of datum represents one of the main obstacles to scalability for many cache coherence protocols. The GLOW extensions are intended for systems with complex networks of interconnected SCI rings, (e.g., large networks of workstations). GLOW extensions are based on building k-ary sharing trees that map well to the underlying topology. In contrast STEM is intended for systems where GLOW is not applicable (e.g., topologies based on centralized switches). STEM defines algorithms to build and maintain binary sharing trees. We show that latencies of GLOW reads and writes grow only logarithmically with the number of nodes sharing, in contrast to SCI where latencies grow linearly, therefore validating GLOW as a good solution to efficient wide sharing of data. Previous work showed the same for STEM