Vorschau

PDF, Englisch Download (836kB) | Nutzungsbedingungen

Abstract

Future data center configurations are driven by total cost of ownership (TCO) for specific performance capabilities. Low-latency interconnects are central to performance, while the use of commodity interconnects is central to cost. This paper reports on an effort to combine a very high-performance, commodity interconnect (HyperTransport) with a high-volume interconnect (Ethernet). Previous approaches to extending Hyper-Transport (HT) over a cluster used custom FPGA cards [5] and proprietary extensions to coherence schemes [22], but these solutions mainly have been adopted for use in research-oriented clusters. The new HyperShare strategy from the HyperTransport Consortium proposes several new ways to create low-cost, commodity clusters that can support scalable high performance computing in either clusters or in the data center. HyperTransport over Ethernet (HToE) is the newest specification in the HyperShare strategy that aims to combine favorable market trends with a highbandwidth and low-latency hardware solution for noncoherent sharing of resources in a cluster. This paper illustrates the motivation behind using 10, 40, or 100 Gigabit Ethernet as an encapsulation layer for Hyper-Transport, the requirements for the HToE specification, and engineering solutions for implementing key portions of the specification.