Managing Oversubscription Risk
Of course, oversubscribing network resources can be risky. At times every user might simultaneously access the service. In such cases, something will have to give. Also, sometimes the usage of service between different users might not be as statistically independent as the service provider would like.
An example of such a usage pattern might be in case of digital TV subscribers at a time when the Soccer World Cup final is on. When this scenario occurs, not only is a larger percentage of users than normal likely to watch TV at the same time, but to make matters worse, they also all want to watch the same program. Digital TV signals are compressed to reduce their bandwidth requirements, taking advantage of the fact that there is generally only a small difference between successive picture frames. However, cuts between scenes lead to temporary bandwidth spikes because an entire new picture needs to be transmitted, not just a delta of it. With different people generally watching different programs, the occurrence of spikes is somewhat distributed, which leads to a statistical effect that can be exploited. However, when everybody watches the same program, everybody requires simultaneously the same spike in bandwidth. This means that picture quality is likely to suffer if the underlying network connections are oversubscribed and not all bandwidth requirements can be satisfied.
Depending on the service, service providers can use a variety of techniques to mitigate those risks; for example:
■ Reprovisioning—In the course of monitoring current service levels and service consumption (see next section), a growing problem related to overusage often can be detected. In such cases, the service provider might be able to temporarily increase available capacity in some circumstances. For example, if core network bandwidth is running low, bandwidth on a network backbone might be increased by temporarily buying additional capacity from another service provider. If in the case of a web-hosting provider server capacity is running low, it might be possible to temporarily bring other servers online.
■ Imposing admission control—A service provider might be able to impose a scheme in which users do not get access to a service if usage is above a certain level and admission of more users would lead to deterioration of everyone's service level. The idea is that sometimes it is better to make one or a few users unhappy instead of everyone. An example is telephone service, in which a user might experience a "busy out" when trying to place a call when a large number of calls already are in progress. The service provider essentially decides to block the call attempt instead of having everyone's call quality deteriorate.
Simple admission-control schemes operate according to a first-come, first-served manner. More sophisticated schemes can make further differentiation based on a policy. For example, they might assign priorities or weights to certain categories of users—a "gold" user gets preferred treatment over a "bronze" user. Ideally, an admission-control scheme takes individual customers' service level objectives into account and gives priority to those customers that the service provider can least afford to disappoint.
In the end, oversubscribing resources in the network is a calculated gamble. Its purpose is to achieve a good balance between utilization of resources (which, if left idling, increase the cost of providing instances of the service) and keeping within agreed service levels (which, if violated, lead to decreased profitability, not just because of the cost that is associated with penalties, but because of lower customer satisfaction and the resulting loss of revenue).
Figure 11-8 depicts this situation. The figure also illustrates how the trade-off can be assessed and hence how the gamble can be calculated:
■ Part (a) of the figure depicts the drop in service level in relation to the degree of oversubscription. By superimposing the service level objective on the curve, the level of oversubscription that is still acceptable can be obtained. (A different interpretation of the same curve that leads to similar conclusions is that the curve depicts the probability that a certain service level will be met at a given degree of oversubscription. The threshold indicates the probability that is still just acceptable.)
■ Part (b) of the figure depicts the increase in utilization that can be achieved by oversubscription. Taking the degree of oversubscription that is still acceptable as was just identified using curve (a), it is possible to derive the level of resource utilization that can be accomplished with this degree of oversubscription. This yields important information about how the network needs to be dimensioned and how to calculate the cost basis for guarantees of a particular service level.
Figure 11-8 Trade-Off Assessment—Acceptable Level of Oversubscription and Achievable Level of Utilization
Figure 11-8 Trade-Off Assessment—Acceptable Level of Oversubscription and Achievable Level of Utilization
(a) Service level (b) Utilization when oversubscribed when oversubscribed
What all of this also implies is that whenever a service level objective really, truly, positively requires an absolute guarantee, not just a probabilistic or statistical one, the required resources need to be dedicated to this particular instance of the service. Of course, this has its price and is a main reason why services with hard-and-fast guarantees tend to be much more expensive than services with softer guarantees.
Continue reading here: Anticipating Problems Before They Occur
Was this article helpful?