A network is congested when one or more network components must discard packets due to lack of buffer space. Given the above architecture, it is possible to see how network congestion can occur. A source of data flow on the network cannot reserve bandwidth across the network to its data’s destination. It, therefore, is unable to determine what rate of data flow can be sustained between it and the destination.
If a source transmits data at a rate too high to be sustained between it and the destination, one or more routers will begin to queue the packets in their buffers. If the queueing continues, the buffers will become full and packets from the source will be discarded, causing data loss. If the source is attempting to guarantee transmission reliability, retransmission of data and increased transmission time between the source and the destination is the result. Figure 2 from [Jain & Ramakrishnan 88] demonstrates the problem of congestion.
As the load (rate of data transmitted) through the network increases, the throughput (rate of data reaching the destination) increases linearly. However, as the load reaches the network’s capacity, the buffers in the routers begin to fill. This increases the response time (time for data to traverse the network between source and destination) and lowers the throughput.
Once the routers’ buffers begin to overflow, packet loss occurs. Increases in load beyond this point increase the probability of packet loss. Under extreme load, response time approaches infinity and the throughput approaches zero; this is the point of congestion collapse. This point is known as the cliff due to the extreme drop in throughput. Figure 2 also shows a plot of power, defined as the ratio of throughput to response time. The power peaks at the knee of the figure.
- Transmission overhead and message size
Figure 3: The transmission overhead against packet size for message length 1000.
We define the transmission overhead to be the number of bits, communicated for a message, that do not represent the data bits of the message. In particular, packet headers and acknowledge packets determine the transmission overhead. Most forms of serial communication use encoding, e.g., adding stop bits, to ensure that the receiving side correctly interprets the serial bit stream. DS links use an encoding scheme that extends each byte, 8 bits, to a 10 bit token. The transmission overhead is therefore at least 20 %.
Figure 2 shows the format of data and acknowledgement packets that we use for the DSNIC protocol. It shows that, apart from the DS link routing header, three characters are used for protocol specific information, independent of the payload size. Together with the acknowledgement scheme, this allows us to calculate the transmission overhead for different packet sizes. Figure 3 shows the transmission overhead against packet size, or, more precisely, against payload size, for sending a 1000 byte message. The packet size strongly influences the required number of packets and packet headers, and thereby the transmission overhead.
Figure 4: The maximum network throughput versus packet size.
Figure 4 shows the influence of the packet size on the maximum network throughput for a 512 end-node Clos network under random traffic. This graph shows an optimal network throughput for packet size 28. For packets smaller than 28, the network throughput drops due to the domination of the transmission overhead. For packets larger than the optimum, the network throughput becomes worse due to network congestion.
The header of each packet needs to be processed by the DSNIC. This processing requires time. Using a small packet size, such as the optimal 28, requires a lot of processing to achieve full bandwidth communication. To keep this processing from becoming the system’s bottleneck, we choose not to fix the packet size, but to make it adaptable so that its influence on the performance of the DSNIC can be investigated. We only support powers of two for the packet size to accommodate the implementation.
|
Other factors that affects the response times on a LAN include; · Speed of devices · Processing time · Priority nodes · Quality of transmission. |
Leave a Reply
You must be logged in to post a comment.