A Canonical Standard

XDR's approach to standardizing data representations is canonical. That is, XDR defines a single byte order, a single floating-point representation (IEEE), and so on. Any program running on any machine can use XDR to create portable data by translating its local representation to the XDR standard representations. Similarly, any program running on any machine can read portable data by translating the XDR standard representations to its local equivalents. The single standard completely decouples programs that create or send portable data from those that use or receive portable data. The advent of a new machine or a new language has no effect upon the community of existing portable-data creators and users. A new machine joins this community by being "taught" how to convert the standard representations and its local representations; the local representations of other machines are irrelevant. Conversely, the local representations of the new machine are also irrelevant to existing programs running on other machines. Such programs can immediately read portable data produced by the new machine because such data conforms to the canonical standards that they already understand.

There are strong precedents for XDR's canonical approach. For example, TCP/IP, UDP/IP, XNS, Ethernet, and, indeed, all protocols below layer five of the ISO model, are canonical protocols. The advantage of any canonical approach is simplicity; in the case of XDR, a single set of conversion routines is written once and is never touched again. The canonical approach has a disadvantage, but it is unimportant in real-world data transfer applications. Suppose two Intel machines are transferring integers according to the XDR standard. The sending machine converts the integers from Intel host byte order to XDR byte order; the receiving machine performs the reverse conversion. Because both machines observe the same byte order, their conversions are unnecessary.

The time spent converting to and from a canonical representation is insignificant, especially in distributed applications. Most of the time required to prepare a data structure for transfer is not spent in conversion but in traversing the elements of the data structure. To transmit a tree, for example, each leaf must be visited and each element in a leaf record must be copied to a buffer and aligned there; storage for the leaf may have to be de-allocated as well. Similarly, to receive a tree, storage must be allocated for each leaf, data must be moved from the buffer to the leaf and properly aligned, and pointers must be constructed to link the leaves together. Every machine pays the cost of traversing and copying data structures whether or not conversion is required. In distributed applications, communications overhead--the time required to move the data down through the sender's protocol layers, across the network and up through the receiver's protocol layers--dwarfs conversion overhead.