1, any single VC can sustain the peak link rate by Equation (2). Thus, link utilization is maximized and transmission time is minimized.

Now consider adaptive credit control. As in the static case, M needs to be large for increased link utilization and reduced transmission time. For adaptive buffer allocation, M needs to be large also for fast ramp up [15] as explained above.

Intuitively, receiver-oriented adaptation needs more buffer than does sender-oriented adaptation, because receiver-oriented adaptation involves an extra round-trip delay for the receiver to inform the sender of the new allocation. Thus the minimum buffer size for receiver-oriented adaptation is increased from RTT to 2*RTT. Suppose that the total memory size is larger than the minimum 2*RTT, e.g., as given by Equation (5). Then the part of the memory that is above the minimum 2*RTT will provide "headroom" for each VC to increase its bandwidth usage under the current buffer allocation. If the VC does increase its bandwidth usage, then the adaptation scheme will notice the increased usage and will subsequently increase the buffer allocation for the VC [12].

The receiver-oriented adaptive buffer allocation scheme in [13] uses M given by Equation (5). Analysis and simulation results have shown that with this choice of M the adaptive scheme gives good performance in utilization, fairness, and ramp up [13].

Link-By-Link Flow Control to Increase Quality of Control

Link-by-link flow control has shorter and more predictable control loop delay than does end-to-end flow control. This implies smaller memory requirements for switching nodes and higher performance in utilization, transmission time, fairness, and so on.

Link-by-link flow control is especially effective for handling transient "cross" traffic. Consider Figure 11, where T is an end-to-end flow-controlled traffic using some end-to-end transport-level protocol such as TCP and X is high-priority cross traffic. If X uses the whole bandwidth of the Switch3's output link, then the entire window of T for coveting the end-to-end round-trip delay would have to be buffered to avoid cell loss. With link-by-link flow control, all the buffers on the path from the source of T to Switch3 can be used to prevent cell loss. In contrast, without link-by-link flow control, only the buffer at the congestion point (i.e., Switch3 in this case) can be used for this purpose. The argument for making efficient use of buffers is similar to that for making efficient use of dams in the flood-control analogy described above.

Figure 11

(a) With link-to-link flow control, all buffers on the path leading to the congestion point (Switch3) where traffic T meets cross traffic X can be used for preventing cell loss; (b) without link-by-link flow control, only the buffer in Switch3 can help.



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement