Cloud Economics

Summary: Cloud Economics

The lecture examines when cloud computing is economically worthwhile rather than assuming that moving to the cloud is always cheaper. Organizations must compare cloud services with owning and operating infrastructure, taking account of demand patterns, operating costs, service requirements, and the risks of provisioning too much or too little capacity.

Why cloud services can be economical

• Pooled, standardized infrastructure: Cloud providers combine resources into shared pools. Because many customers’ workloads are served by the same infrastructure, providers can use capacity more efficiently than an organization provisioning for its own peak demand.
• Statistical multiplexing: An organization may buy enough systems for a rare busy period, leaving them underused much of the time. A shared pool can serve different workloads as their demands rise and fall, potentially increasing utilization and lowering the cost per unit of service.
• Economies of scale: Large-scale operations may reduce overheads such as power, cooling, maintenance, equipment support, and staffing. Hardware also becomes obsolete, often within a few years, making the cost of continually maintaining in-house infrastructure important.
• On-demand resources and utility pricing: Cloud capacity can be provisioned or released as needed, and customers typically pay for what they use. This can reduce the need to buy capacity in advance, provided resource changes are timely and do not require costly manual management.
• Location independence: Users can access applications and computing resources over a network instead of travelling to or relying on a particular machine. This can improve convenience and user experience, but depends on network performance and connectivity.

Demand variation and capacity planning

A fixed facility serving highly variable demand may achieve lower utilization than one serving a smoother workload. Combining multiple demand sources can smooth overall demand: for n independent, similar workloads, the coefficient of variation (CV) of their combined demand falls roughly in proportion to 1/√n.

The CV is the standard deviation divided by the mean. It helps compare the relative variability of demand, even when two workloads have different average sizes. In this setting, smoother demand generally makes capacity planning easier.

The smoothing benefit depends on how workloads relate to one another:

• Independent or complementary workloads—for example, customers that are active at different times—can help smooth demand.
• Perfectly correlated workloads, which peak together, do not provide the same smoothing benefit.
• Choosing customer or workload segments with different demand patterns can therefore matter to a provider’s economics.

Statistical multiplexing can reduce the need for excess capacity and the risk of unmet demand, lost revenue, or service-level agreement (SLA) penalties. However, congestion can still occur if many users demand resources at once.

Location, latency, and architecture

Cloud access relies on network connectivity, so the distance to services and the number of network hops can affect latency. Latency matters to user experience, especially for interactive or voice applications. Supporting users around the world may require a geographically distributed service architecture.

Network design also brings costs and technical considerations, including fixed and usage-based network costs, interoperability overhead, reliability, and availability. These factors can affect the overall economic case for cloud services.

A break-even comparison for cloud versus owned infrastructure

The lecture presents a simplified way to compare costs over a period of time:

• Let D(t) be resource demand over time.
• Let P be peak demand and A be average demand.
• Let B be the cost per unit of owned, baseline capacity.
• Let C be the cloud cost per unit.
• Define the utility premium as U = C/B, the cloud unit cost relative to the owned unit cost.

Under this simplified model:

• Cloud cost over the period is based on the amount of capacity actually used: C × ∫D(t)dt, or U × B × A × T when expressed using average demand over period T.
• Owned infrastructure must be sized for peak demand, giving a cost of approximately P × B × T.
• On these assumptions, cloud is cheaper when P/A > U: the peak-to-average demand ratio exceeds the cloud utility premium.

This is only a simplified comparison. A cloud service need not have a lower unit price to be economical; the result depends on demand patterns and the costs of ownership and operation.

On-demand provisioning and the cost of mismatch

Owning a fixed amount of capacity creates a penalty when it does not match instantaneous demand:

• If capacity exceeds demand, the organization pays for unused resources.
• If demand exceeds capacity, the organization may miss service or incur other delivery penalties.
• The lecture describes this penalty as relating to the gap between demand and available resources over time.
• Flat demand is easier to serve with fixed capacity. Gradually changing demand may be handled with periodic provisioning, but rapidly changing or nonlinear demand is more difficult.
• With exponentially increasing demand, provisioning delays can cause the gap between demand and capacity—and the resulting penalty—to grow rapidly. Demand patterns and provisioning delays must therefore be understood before choosing a capacity strategy.

Assignment mentioned

The instructor gives an exercise for later discussion:

• Consider an organization with a peak computing demand of 120 units.
• Use the specified demand-over-time function and cloud resource-provisioning rule from the assignment.
• Account for a provisioning delay, Δt, when adding on-demand resources.
• Assume the cost of provisioning one cloud resource unit for one unit of time is 0.9 units.
• Calculate any resulting penalty and explain what the result indicates about provisioning delay and service delivery.

The lecture concludes that cloud economics depends on the relationship between demand and capacity, the amount of demand smoothing possible, the costs of operating infrastructure, and the ability to provision resources in time.

Speakers and sources

• Speaker: One unnamed instructor or lecturer.
• Other featured speakers or cited sources: None identified in the subtitles. Examples such as a university computer lab, network equipment, and car rental are used as illustrations.