Cloud Cost Optimization

Instance types are optimized for memory, database, computing, graphics, storage capacity, throughput, etc. The right choice depends on the nature of your application.

Monitoring CPU and memory usage over time gives a clear picture of the instances that can be cut to size avoiding wastage. Releasing what you don’t need can save you a lot of money.

Tracking utilization using built-in tools and reevaluating your requirements in the light of the data it yields is important. By making this an ongoing process, you can have a better grasp of the actual processing and memory your cloud applications require. Consolidating applications on low-CPU-utilising instances can also help reduce costs.

There are geographical considerations as well—where the users largely come from and what the cost is in each region. If there is a geographical concentration of users in a particular region, optimization by way of allocating resources in that region can be opted.

For workloads that demand resources for select periods of time (such as batch processing of data) spot instances created at the time of demand are a more economical option. Opportunistically turning instances on and off based on actual usage is a great way to save up to 75% on costs. Automated scheduling makes this more efficient. Cloud platforms can be configured to auto-scale resources based on a set of rules.

Re-architecting the application to leverage spot instances, which can be terminated anytime, is also a cost-cutting measure. However, spot instances are not suitable for stateful applications, databases, etc.

For predictable or unchanging workloads, Reserved Instances (RI), with their guaranteed capacity, are a better choice. RI comes with significant discounts in return for a long-term commitment. It is just as important though to ensure that RI discounts are always availed of and not wasted.

The selection of GPU over CPU is another decision that has cost implications. GPUs are preferred over CPUs if the applications involve large computations (such as machine learning and 3D rendering). Appropriate CPU or GPU instances can be created depending on the requirement.

Diverse offerings such as Amazon EC2 CPU instance, dedicated GPU instances such as G3 and the less expensive Elastic Graphics can add GPU to existing CPU instances. For machine learning, there are higher-value EC2 - P2 instances with support for parallelism and the less expensive Elastic Inference that can be attached to a regular EC2 instance.

Moving data in and out of storage is expensive, so the choice of storage must be decided at the outset and must be guided by your requirements. Amazon S3 Standard is an object store for data that needs to be retrieved often. Amazon S3 Standard - IA and related variants are better for backing up data that is not likely to be accessed for a very long time. For compliance backups and archival storage, Amazon Glacier is a cheaper choice. You can also reduce the costs of data access by selectively retrieving objects from S3 using SQL expressions, thereby reducing transfer and compute costs.

Performance requirements can be better met by storage options such as block storage (for example, Amazon Elastic Block Store) or scalable Linux file system (for example, Amazon Elastic File System). There are other specialized systems for requirements such as high-performance computing and Windows native applications with trade-offs among throughput, retrieval time, and cost.

If access becomes infrequent over time, you can also make use of lifecycle policies to move data from one storage type to another. By implementing lifecycle policies and automatic rules to delete or move data once its lifetime has been served, you can get rid of unwanted data and storage costs. Dated snapshot is a common source of storage wastage, which can be eliminated using the appropriate policy and lifecycle management tool.

The licensing cost of the base operating system can spike the total cost of running an application so it needs to be managed carefully. An instance of a Linux system like Debian or Ubuntu on a virtual machine costs less than Windows or a supported RHEL instance.

To run .NET applications, you need to pay for Windows license. If they are migrated to .net core framework, they can also run on Linux instances that have low or no license fees. Thus migration to .net core can be a cost reduction strategy. If you have already purchased Windows license, you might want to consider taking dedicated hosts to take advantage of the existing licenses.

Avoiding dependence on proprietary databases like Oracle and MS SQL can save money too. Using tools like AWS Schema Conversion, your data can be moved to a less expensive database. Even if you must use a proprietary system, a careful selection of the version, such as MS SQL Web / Standard / Enterprise, can help you realize some cost savings.

Several imperatives push enterprises towards a multi-cloud model—varying department needs, pricing schemes, SLAs, security, and so on. But managing the diverse products from multiple vendors requires specialized tools and additional training, which makes it cost-intensive. A single cloud, on the other hand, eliminates the administrative hassle of switching between platforms and paying for network traffic between clouds.

The chances of getting volume discounts increase when you deploy all your resources on a single cloud. While single cloud is advisable for many reasons, there can be limits to such benefits. The right choice can be made only after careful consideration of your application needs and environment.