Software-Defined Cluster

The cluster architecture has played an important role in high-end computing for the past 20 years. With the advent of Internet services, big data, and cloud computing, traditional clusters face three challenges： 1） providing flexible system balance among computing, memory, and I/O capabilities; 2） reducing resource pooling overheads; and 3） addressing low performance-power efficiency. This position paper proposes a software-defined cluster （SDC） architecture to deal with these challenges. The SDC architecture inherits two features of traditional cluster： its architecture is multicomputer and it has loosely-coupled interconnect. SDC provides two new mechanisms： global I/O space （GIO） and hardware-supported native access （HNA） to remote devices. Application software can define a virtual cluster best suited to its needs from resources pools provided by a physical cluster, and traditional cluster ecosystems need no modification. We also discuss a prototype design and implementation of a 32-processor cloud server utilizing the SDC architecture.