Determining the Winner Set

For each prefix, the CNA system computes the probability of the link being removed from winner-set contention due to load optimization. This probability factor is computed each time a load sample is received from the link-monitor process.

The probability factor is derived from two sources:

• current load, as reported by the link-monitor process
• usage thresholds which you configure in a table, using the load threshold-table command

Each row is also identified by a usage level (expressed in Mpbs) and each cell in the table contains a user-defined threshold which determines when the link will be removed from the winner set. Though only one value is displayed in the example table, this threshold is actually a range of values--as long as usage remains below the lower number, the probability of the link being removed is 0; when usage exceeds the larger number the probability is 1 and the link is automatically removed.

The higher number, the max-avoidance value, is specified in megabits per second (Mbps), and is the number displayed in the example table.

The lower, start-avoidance number, is expressed as a percentage to be subtracted from the larger number. The range is expressed as start-avoidance: max-avoidance, with the start-avoidance value implied (not stated explicitly) as long as it is the default value.

The maximum threshold must be explicitly configured, using the link command in threshold configuration mode.

The minimum threshold defaults to 10 percent of the usage level value from the level the cell is in, divided by the number of cells. This is so the same amount of tolerance is assigned to each link at each level, irrespective of the threshold value. This can be changed globally by using the start-avoidance percent-below-max command, or per cell using the start-avoidance percent-below-max parameter of the link command. In the latter case, the value corresponds to a percentage less than your configured maximum value.However, if you change the percentage value, you should watch the results carefully to be sure it achieves the intended results.

When actual total usage across all links exceeds the number specified, the next row in the table will be used. The last row should contain the maximum usage allowed by your service providers.

A configured table which does not change the minimum threshold percentage might look like Table 4:

As long as total usage across all three links is less than 90 Mbps, the start-avoidance value would be 27 Mbps (10 percent less than the maximum threshold for the link, which is 30 Mbps) for each of the three links. As long as usage on a link remains below this level, the probability of it being removed from the winner set is 0.

When the decision-maker process reads the table, it will seek the level with the smallest total load at which the actual load sample obtained from the edge router is less than the sum of the configured maximum avoidance values.

For example, if total usage across all links were 110 mbps, the decision-maker process would use the thresholds in the third row in the example table. That is the level with the smallest total load where the actual sample (110) is less than the configured total (118).

If total usage across all links is less than 90 mbps, the first row of the table would be used.

If total usage across all links is 90 or higher but less than 104, the thresholds in the second row would be used.

When usage is 104 or higher but less than 118, the thresholds in the third row will be used.

The thresholds in the fourth row will be used when usage reaches 118 but less than 132, which is the maximum bandwidth you are entitled to use according to your service provider contracts.

Probability will be 1 if current usage is greater than the specified maximum threshold.

If current load is greater than the beginning threshold (start-avoidance) but less than or equal to the maximum threshold (max-avoidance), then the probability of the link being removed from winner-set contention because of load will be somewhere between 0 and 1.

As long as the probability is 0, the link will automatically remain a candidate for the winner set for the prefix. As the number rises, the probability of the link being removed from the winner set increases until probability reaches 1, at which time the link will be automatically removed from the winner set. A prefix will evaluate these probabilities every time it is processed by the decision-maker process and the evaluation is allowed (see Specifying the Evaluation Interval).

The decision-maker process will use this probability to decide whether the link should remain in the winner set, and a winner will be chosen according to the following set of rules:

• If a winner can be selected from the regular winner set, the process ends for this prefix and the threshold-table level indicator in the show prefix command output will be r (see the section Show Prefix Command in Chapter 3: Adaptive Path Control).
• If all links have been removed from the winner set before a winner can be selected, the decision-maker process will broaden its selection range by using the extended winner set value. If a winner can be selected from the extended winner set, the process ends and the threshold-table level indicator will be e.
• If there are no links in the extended winner set, the decision-maker process will jump to the next level in the table and repeat the process. The level indicator will be j. This is a signal that winner set is too narrow and you should consider increasing the size of the extended winner set value.
• If at any time all three links are considered to be hot, the process will be repeated for the links whose probability is not equal to one, until a winner is selected. The level indicator will be h.
• If the decision-maker process is not able to determine a winner by this process because total usage has exceeded the highest load defined by your table, then the route will be determined by simple load balancing. This indicates that either your table is misconfigured or you have insufficient capacity to handle your traffic.

This entire process is repeated several times a second, individually for each prefix that moves through the decision-maker processing queue. Over time, the distribution of all prefixes will be spread among your service providers in the most cost-effective manner (as defined by the criteria you input into the configuration).