Load balancing

Both performance properties and weighted load balancing properties can split, or balance, traffic load between two or more data centers.

These are the primary types of GTM load balancing. In addition, Performance with load feedback, has three variants.

  • Weighted random load balancing

  • Performance-based load balancing

  • Performance-based load balancing with load feedback.

    • With XML load objects

    • With non-XML load objects

    • With download score based load balancing

Load balancing compared to end-user performance

Load balancing is often in tension with end-user performance. A very simple example: suppose that you have two data centers, one in New York and one in Singapore, and that your users are financial industry workers in the U.S. and in Asia. Suppose further that you have configured a 50/50 split of traffic between the two data centers. During business hours in New York, almost all of your traffic is coming from the U.S., as your Asian users are probably asleep. If GTM actually sent half of that traffic to Singapore, many of the U.S. users would receive poor performance. So you can see that there is a trade off to be made between performance and load balancing. How you control this trade off varies depending on the type of load balancing you choose. In some modes, this is controlled by a number you configure called the load imbalance factor (LIF).

Load imbalance factor

The load imbalance factor (LIF) controls how imbalanced GTM allows the load to be; the factor by which the demand sent to a data center is permitted to exceed the configured value. For example, with a data center traffic allocation of 25 percent and a LIF of 50%, the demand sent to the data center is allowed to grow to 37.5 percent (25% + (25% * 50%) ) before the load balancer starts shifting load away from it.

📘

This option applies only to the following property types.

  • Performance-Based Load Balancing

  • Performance-Based Load Balancing with Load Feedback (targets computed from configured weights)

  • Performance-Based Load Balancing with Load Feedback Based on Liveness Test Download Scores

You can use the LIF with either percent-based load balancing without load feedback, or with one of the load feedback modes that computes targets based on percentages. The LIF is not operational when using XML load objects to report current, target, and maximum loads for a datacenter. In non-load feedback mode, the imbalance factor simply raises the allowed traffic percentage sent to each data center by the requested amount. In load feedback mode with automatically computed targets, the LIF raises the targets after they are computed.

Load imbalance is not necessarily bad. The example in Load Balancing Compared to End User Performance shows clearly that good end-user performance often requires an imbalanced load. For good performance, you usually need to allow a certain amount of imbalance. The load imbalance factor controls how much imbalance you will permit. Reducing the LIF to below 10% is likely to result in oscillation, meaning the load shifts back and forth between the two data centers.

The default load imbalance factor is 10%.

Weighted random load balancing

Weighted random load balancing completely ignores performance. It therefore gives very accurate load balancing, but at a cost of potentially sending some users to a data center that is far from the optimal choice for performance. If precise load balancing is more important than any other consideration, weighted load balancing is a good choice.

Because the answers returned by weighted random load balancing are computed randomly, without regard to the identity of the requester, it is also a good choice if you need to balance traffic that is generated by only a small number of client nameservers.

Weighted random load balancing does not use nameserver demand estimates.


Did this page help you?