Network protocol design principles

Protocol Layering Usually, protocols are layered. For example, one layer might describe how to encode text (with ASCII, say), while another describes how to inquire for messages (with the Internet's simple mail transfer protocol, for example), while another may correct errors (with the Internet's transmission control protocol), another handles addressing (say with IP, the internet protocol), another handles the error detection (with the internet's point-to-point protocol), and another handles the physical form of the bits, (with a V.42 modem, for example).

Layering allows the parts of a protocol to be designed and tested without a combinatorial explosion of cases, keeping each design relatively simple. Layering also permits familiar protocols to be adapted to unusual circumstances. For example, the mail protocol above can be adapted to send messages to aircraft. Just change the V.42 modem protocol to the INMARS LAPD data protocol used by the international marine radio satellites.

Error Detection[?] and Correction It's a truism[?] that communication media are always faulty. The conventional measure of quality is the number of failed bits per bit transmitted. This has the wonderful feature of being a dimensionless figure of merit that can be compared across any speed or type of communication media.

In telephony, failure rates of 1x10^-4 bit per bit are faulty (they interfere with telephone conversations), while 1x10^-5 bit per bit or more should be dealt with by routine maintenance (they can be heard).

Communication systems correct errors by selectively resending bad parts of a message. For example, in TCP (the internet's Transmission Control Protocol), messages are divided into packets, each of which has a checksum. When a checksum is bad, the packet is discarded. When a packet is lost, the receiver acknowledges all of the packets up to, but not including the failed packet. Eventually, the sender sees that too much time has elapsed without an acknowledgement, so it resends all of the packets that have not been acknowledged. At the same time, the sender backs off its rate of sending, in case the packet loss was caused by saturation of the path between sender and receiver. (Note: this is an over-simplification: see TCP for more detail)

In general, the performance of TCP is severely degraded in conditions of high packet loss (more than 0.1%), due to the need to resend packets repeatedly. For this reason, TCP/IP connections are typically either run on highly reliable fiber networks, or over a lower-level protocol with added error-detection and correction features (such as modem links with ARQ). These connections typically have uncorrected bit error rates of 1x10^-9 to 1x10^-12, ensuring high TCP/IP performance.

Resiliency Another form of network failure is topological failure, which a communications link is cut. Most modern communication protocols periodically send messages to test a link. In phones, a framing bit is sent every 24 bits on T1 lines. In phone systems, when "sync is lost", fail-safe mechanisms reroute the signals around the failing equipment.

In packet switched networks, the equivalent functions are performed using router update messages to detect loss of connectivity.

See Also

• Radia Perlman[?], Interconnections: Bridges, Routers, Switches, and Internetworking Protocols (2nd Edition). Addison-Wesley 1999. ISBN 0201634481. In particular Ch. 18 on "network design folklore".
• Gerard J. Holzmann[?], Design and Validation of Computer Protocols. Prentice Hall, 1991. ISBN 0-13-539925-4. Also available online at http://spinroot.com/spin/Doc/Book91