Figure 2: By analyzing usage behavior, cloud resources can be scarcely commissioned in order to cover possible load peaks via elasticity.

© Source: Apptio

The autoscaling function can be used to automatically adjust the required cloud capacities based on predefined rules. If there is little computing to be done, low computing capacities are sufficient; if computing-intensive tasks are pending, capacity is increased on-the-fly. An autoscaling group defines the minimum and maximum capacity to be made available, which instance types are to be used and how scaling is to take place.

Scaling can be manual, time-controlled or automatic based on application-specific metrics (monitoring) and/or predictions. This ensures the availability of services and reliable performance levels in a demand-oriented and cost-efficient manner.

Although modern architectures usually allow autoscaling, it must be clarified in each individual case which applications are suitable for this. The application must at least support load balancing so that the load can be distributed across several compute instances. Applications that show a stable and recurring pattern in the load volume (depending on the time, day of the week and/or calendar week) are very good candidates for autoscaling. Auto-scaling groups can be populated with both on-demand resources and spot instances, as explained in the next section.

5. spot instances - worthwhile additional effort

Spot instances offer a particularly attractive cost structure. However, they are not guaranteed: The provider can terminate the instance at any time if they need it to back up a workload elsewhere, for example. Typically, there is a two-minute warning before the instance is shut down. It's a bet on free capacity in the cloud, so to speak.

One thing is clear: this is not suitable for every application! What are spot instances suitable for?

Computing-intensive tasks that can be interrupted and resumed at any time: Batch jobs, compiler runs or learning processes for artificial intelligence, complex video or image editing processes, scientific calculations or simulations, computationally intensive financial analyses and models.

• Distributed databases that store data even when individual instances are restarted.
• Big data: All types of mass data processing.
• Testing: Load tests, regression tests, security tests or tests with mass data.

The reward for the effort: spot instances are available for up to a tenth of the price of on-demand resources. Tools that visualize cloud usage in terms of spot, reserved instances, savings plan and on-demand usage are therefore worthwhile. They help to understand which workloads are suitable and how the spot instances, which providers make available according to different regulations, can best be used. However, a system for managing interrupted spot instance workloads is also important.

6. waste management - clean out, but with caution

Figure 3: Practical rightsizing: Automated determination of booked cloud services based on cloud billing data compared to monitoring data makes it clear what was oversized or not used at all.

© Source: Apptio

Waste management is about identifying unused cloud resources and avoiding unnecessary expenditure. Sounds simple at first - but with typically several hundred thousand cloud resources, it is not possible to find out manually which resources are actually not needed.

By comparing billing data with monitoring data, tools can identify potentially superfluous cloud expenditure. In addition, among all the hundreds of thousands of cloud resources, those that are not needed at night or at the weekend can be identified. A distinction can also be made between productive and non-productive environments.

But beware: unused or barely used cloud instances or cloud services should not simply be switched off automatically, but notes on this should be created as tickets in the appropriate Jira projects. Only the cloud architects and DevOps teams can make informed decisions about which cloud services are actually superfluous and which are not. Even if monitoring shows that they are hardly used, there is a risk of production downtime if individual cloud instances or services are not available without detailed knowledge of the context and architecture.

7 Workload management - creating comparability

Figure 4: By comparing billing data with monitoring data, Apptio tools identify potentially superfluous cloud expenditure. This helps cloud architects to identify services that can be switched off from thousands of cloud resources.

© Source: Apptio

The costs for virtual machines vary greatly from provider to provider. The designation of the workloads in the cloud, the offers and configurations of the cloud providers also vary significantly. This makes it difficult to compare offers. However, the price of workloads is not the only reason to choose a provider. Cloud providers try to bind customers to their cloud platform by offering ready-made cloud services. And their use is indeed attractive, as the provider takes care of maintenance and updates.
As a decision-making aid, tools can show which cloud instances are most likely to cover a particular workload. In most cases, there are between 40 and 50 similar configurations that could be suitable in principle. If the tool knows the configurations of existing on-premises servers, it can also determine which cloud instances could replace them. This enables comparisons to be made as to which cloud offerings are best suited - also, but not only, in terms of the price/performance ratio.

Transparency creates responsible handling

The author: Thomas Köppner is a Solution Consultant at Apptio.

© Apptio

The important thing is: In addition to all cloud cost optimizations, a responsible use of resources in the cloud is required. The first step is to create the necessary sense of responsibility through cost transparency and allocation according to who is responsible. The teams need to understand what costs they are incurring through their cloud usage and what the core cost drivers are. In the second step, it makes sense not only to show the costs, but also to charge them to the teams via internal cost allocation. If cost-effective use of the cloud is one of the objectives, cloud costs will no longer rise unchecked, but will be invested in cloud services where the business benefits justify the expenditure.