4. TRANSPORTS AND SEMANTICS

The RPC protocol can be implemented on several different transport protocols. The RPC protocol does not care how a message is passed from one process to another, but only with specification and interpretation of messages. However, the application may wish to obtain information about (and perhaps control over) the transport layer through an interface not specified in this document. For example, the transport protocol may impose a restriction on the maximum size of RPC messages, or it may be stream-oriented like TCP with no size limit. The client and server must agree on their transport protocol choices.

It is important to point out that RPC does not try to implement any kind of reliability and that the application may need to be aware of the type of transport protocol underneath RPC. If it knows it is running on top of a reliable transport such as TCP [6], then most of the work is already done for it. On the other hand, if it is running on top of an unreliable transport such as UDP [7], it must implement its own time-out, retransmission, and duplicate detection policies as the RPC protocol does not provide these services.

Because of transport independence, the RPC protocol does not attach specific semantics to the remote procedures or their execution requirements. Semantics can be inferred from (but should be explicitly specified by) the underlying transport protocol. For example, consider RPC running on top of an unreliable transport such as UDP. If an application retransmits RPC call messages after time- outs, and does not receive a reply, it cannot infer anything about the number of times the procedure was executed. If it does receive a reply, then it can infer that the procedure was executed at least once.

A server may wish to remember previously granted requests from a client and not regrant them in order to insure some degree of execute-at-most-once semantics. A server can do this by taking advantage of the transaction ID that is packaged with every RPC message. The main use of this transaction ID is by the client RPC entity in matching replies to calls. However, a client application may choose to reuse its previous transaction ID when retransmitting a call. The server may choose to remember this ID after executing a call and not execute calls with the same ID in order to achieve some degree of execute-at-most-once semantics. The server is not allowed to examine this ID in any other way except as a test for equality.

On the other hand, if using a "reliable" transport such as TCP, the application can infer from a reply message that the procedure was executed exactly once, but if it receives no reply message, it cannot assume that the remote procedure was not executed. Note that even if a connection-oriented protocol like TCP is used, an application still needs time-outs and reconnection to handle server crashes.

There are other possibilities for transports besides datagram- or connection-oriented protocols. For example, a request-reply protocol such as VMTP [2] is perhaps a natural transport for RPC. ONC RPC uses both TCP and UDP transport protocols. Section 10 (RECORD MARKING STANDARD) describes the mechanism employed by ONC RPC to utilize a connection-oriented, stream-oriented transport such as TCP.