System area networks (SANs) represent an area of computer architecture that has evolved quickly. The term SAN in this section refers to “system” (not “storage”) area networks. After various competing standardization efforts starting in the late 1990s, the state of the SAN field became temporarily unclear. However, the technology has emerged with a richer set of
features that promise to impact the server and clustering arena.
A SAN uses high-speed connections to attach high-performance computers in a cluster configuration. The configuration delivers very high bandwidth of 1+ GB/sec with very low latency. They are switched, with a typical hub What’s Next 579 supporting 4 to 8 nodes. Larger SANs are built with cascading hubs with cable length limitations that vary from a few meters to a few kilometers.
Interconnections in a SAN differ from other existing high-performance media (such as gigabit Ethernet and ATM) in several ways. SAN adapters implement reliable transport services that are similar to TCP or SPX, but directly in hardware. SANs have very low error rates. SANs are often made highly available by deploying redundant interconnect fabrics.
SANs provide bulk data transfer through a remote direct memory access (RDMA) mechanism. The performance within a SAN resembles more that of a memory subsystem than a traditional network (such as an Ethernet LAN). The initiator specifies a buffer on the local system and a buffer on the remote system. Data is then transferred directly between the local and remote systems by the network adapters without involving either of the host CPUs. Both read and write operations are supported in this manner.