ATM overhead and packet latency tradeoffs
 
|
|||
 | |||
|
Sign up to receive this and other networking newsletters in your inbox.
One of the arguments why packet network technology and fiber ATM alternatives such as Packet over SONET are more viable than cell network technology is that ATM cell overhead dramatically reduces bandwidth efficiency. The real challenge for packet networks is not that latency is unpredictable, but that it dramatically increases the response time of transaction-based application throughput.
Similar to a packet network technology such as Fast Ethernet, Packet Over SONET is a transport technology that encapsulates packet data such as an IP datagram directly into SONET, eliminating ATM's cell overhead. As the argument goes, smaller size packets suffer a large amount of ATM overhead. A 64-byte IP packet will not fit into the 48-byte payload of one ATM cell, so two are needed, with the remainder of the payload of the second cell consisting of padding-bytes that don't contain any data to simply fill out the cell. These two cells will consume 106-bytes, approximately 60% efficiency (64-bytes divided by 106-bytes). and with Packet Over SONET, the overhead is next to nil-only a few addressing bytes per packet, so efficiency is nearly 100%.
Things do get slightly better with larger frame sizes. For a 1,500-byte IP packet, efficiency is about 88%. Padding has much less of an effect on large packets than small packets, since the most a cell would need to be padded is 47 bytes, and 47 compared to 1,500 is small. Cell overhead, having the same amount of impact regardless of frame size, takes up about 10% of bandwidth.
The well known tradeoff is that packet network technologies including Packet Over SONET have absolutely no way of maintaining latency and response time since traffic from any one stream on a packet network completely consumes the media for at least a packet transmission time.
Let's say in order to maintain a certain response time for a 100-byte transaction-base stream, a packet must go out at exactly T=0. In a packet network, this packet must wait at best 0 byte times (the time it takes to transmit a byte of a certain medium, 0.05 microseconds for 155 ATM), meaning there is no packet being transmitted from any other traffic stream. At worst, the packet must wait up to 1,500 byte times (an entire 1,500 byte packet) if another packet begins to be transmitted an instant before T=0. With ATM, both packets can be transmitted at relatively the same time, regardless of when each arrives at the transmission queue because cells from each packet are interleaved. So there is at most a 53-byte wait time, and application traffic that uses small packets-like our transaction based stream will suffer minimal degradation.
So the bottom line is, traffic streams that use small packets need ATM's inherent packet interleaving to keep response time low, even though overall bandwidth efficiency may be lower on a packet network. Streams using larger packets such as file transfers don't need ATM's cell interleaving and consequently aren't as affected by cell overhead. When choosing a transport technology, see how your applications use frame sizes and determine you applications need for low response time.
RELATED LINKS
Cisco's Packet over SONET/SDH (POS) Technology Support; Mission Accomplished
White paper from Cisco.
IP over SONET: An Overview of Enabling Technologies
Fom the University of British Columbia.
PPP over SONET/SDH
An Internet draft from the Internet Engineering Task Force.
