Design and Implementation of a Multi-purpose Cluster System Network Interface Unit Design and Implementation of a Multi-purpose Cluster System Network Interface Unit Design and Implementation of a Multi-purpose Cluster System Network Interface Unit

Today, the interface between a high speed network and a high performance computation node is the least mature hardware technology in scalable general purpose cluster computing. Currently, the one-interface-ts-all philosophy prevails. This approach performs poorly in some cases because of the complexity of modern memory hierarchy and the wide range of communication sizes and patterns. Today's message passing NIU's are also unable to utilize the best data transfer and coordination mechanisms due to poor integration into the computation node's memory hierarchy. These shortcomings unnecessarily constrain the performance of cluster systems. Our thesis is that a cluster system NIU should support multiple communication interfaces layered on a virtual message queue substrate in order to streamline data movement both within each node as well as between nodes. The NIU should be tightly integrated into the computation node's memory hierarchy via the cache-coherent snoopy system bus so as to gain access to a rich set of data movement operations. We further propose to achieve the goal of a large set of high performance communication functions with a hybrid NIU micro-architecture that combines custom hardware building blocks with an o-the-shelf embedded processor. These ideas are tested through the design and implementation of the StarT-Voyager NES, an NIU used to connect a cluster of commercial PowerPC based SMP's. Our prototype demonstrates that it is feasible to implement a multi-interface NIU at reasonable hardware cost. This is achieved by reusing a set of basic hardware building blocks and adopting a layered architecture that separates protected network sharing from software visible communication interfaces. Through diierent mechanisms, our 35MHz NIU (140MHz processor core) can deliver very low latency for very short messages (under 2s), very high bandwidth for multi-kilobyte block transfers (167 MBytes/s bi-directional bandwidth), and very low processor overhead for multi-cast communication (each additional destination after the rst incurs 10 processor clocks). We introduce the novel idea of supporting a large number of virtual message queues through a combination of hardware Resident message queues and rmware emulated Non-resident message queues. By using the Resident queues as rmware controlled caches, our implementation delivers hardware speed on the average while providing graceful degradation in a low cost implementation. Finally, we also demonstrate that an oo-the-shelf embedded processor complements custom hardware in the NIU, with the former providing exibility and the latter performance. We identify the interface between the embedded processor and custom hardware as a critical design component and …