PRIMECLUSTER supports either the Ethernet protocol or IP as the transport layer for the cluster interconnects. This does not mean that TCP/IP is the underlying protocol used by the cluster services.
TCP/IP is designed to reliably deliver messages through a variety of obstacles such as different network types, multiple routers or switches, and failures in the network. TCP/IP works well and has been almost universally adopted as the protocol stack of choice for a wide variety of services. But TCP/IP is not the right choice for cluster interconnects.
The ICF (Internode Communication Facility) protocol used by PRIMECLUSTER is designed specifically for cluster communications. It is a low-latency protocol that guarantees ordered delivery of messages. ICF has a lower overhead than TCP/IP, which is important for parallel applications. ICF uses the Ethernet protocol or is a service over IP (CF/IP).
Note
Devices that only accept TCP/IP cannot route the ICF protocol when it is configured directly on the Ethernet. In this instance, if you need to use a router, it must be a level-two router.
ICF is only usable by the CF internal components and is not available to user-level resources. To provide applications with an access to the cluster interconnect, the Cluster Interconnect Protocol (CIP) is used. CIP provides a standard TCP/IP protocol suite over ICF.
