The InfiniBand structure is a channel-based serial interface instead of public bus, thus separating the I/O subsystem from the CPU/memory. All systems and nodes can be logically connected to the fabric by channel adapters, which can be hosts, adapters (HCAs) or target adapters (TCAs), and also include InfiniBand switches and router extensions to meet the growing demand.
InfiniBand is also a layered protocol (similar to TCP/IP protocol), where each layer is responsible for different functions. The lower layers serve the upper layers, and the different layers are independent of each other, with each layer providing the functions accordingly. infiniBand protocol can meet a variety of different needs, including multicast, partitioning, IP compatibility, flow control, and rate control.
InfiniBand network routing algorithms include shortest path algorithm, Min Hop based UPDN algorithm and fat tree based networking fat tree algorithm.
The algorithms also determine the InfiniBand network topology to some extent, especially in high performance computing and large cluster systems. The topology between the networks must be considered, and the network uplink and downlink blocking conditions also determine the overall network performance. Because tree topologies are clear, easy to build and manage, fat tree network topologies are often used to take advantage of InfiniBand networks and are often used in applications with no or very low blocking rates.