Optimizing Cloud Spend for Cost Efficiency

Finance and engineering worked together to create a cloud spend management roadmap, forming a Central team with engineers from multiple organizations to drive on-demand cost optimization.

The focus was on three things.

• Budget allocation and compliance
• Setting savings itself as a goal
• Securing reliability, sustainability, and control in line with the company's leadership principle of Hate Waste

Initially, engineering teams didn't fully understand cloud usage and efficiency, and finance also had difficulty identifying each team's spending items and how to reduce them. Leadership connected the right people and analytics systems to control the cloud's variable cost structure, and the central team helped each domain team find cost optimization opportunities while maintaining their actual usage responsibilities.

To do this, they built custom dashboards processing Amazon CloudWatch data through Amazon Athena, along with a BI dashboard processing AWS CUR (Cost & Usage Reports). Finance pushed the importance of monthly and quarterly budget management, supporting spend control by team.

The savings strategy was broadly divided into two axes.

• Spending Less: Automatically spinning up AWS resources only when needed in non-production environments, achieving 25% cost savings
• Paying Less: Directly eliminating idle and underutilized EC2 resources based on usage patterns

Specifically, several optimization efforts were carried out. First, all instances were upgraded to the latest generation, and after reviewing both AMD and ARM architectures, internal product lines were moved to AMD CPUs, achieving a 20% price-performance improvement over the previous generation.

For EMR, while continuing to use Spot Instances, instance fleets were introduced to solve the problem of securing capacity during peak hours. This upgrade achieved a 25% reduction in total EMR costs.

For storage, EBS and S3 were the main focus.

• Amazon EBS: After thorough internal validation that the next-generation GP3 met performance requirements, it was first tested in development accounts. As a result, they were able to migrate 500-1000 live volumes in parallel without noticeable impact.
• Amazon S3: 50+PB was moved to Intelligent-Tiering. However, when average object size is too small and the number of objects reaches billions, costs can actually increase, so in such cases policies had to be fine-tuned using S3 lifecycle filters.

These optimizations reduced 2021 AWS On-Demand costs by millions of dollars. However, since much of the work still relied on manual effort, the next step was to introduce more sophisticated analytics tools and automated monitoring to strengthen the Cloud FinOps culture.