Reducing costs up to 20% and raising peak load stability using AWS Graviton with Flipster

Flipster has 1 million registered users and a daily trading volume of around 1 billion US dollars. The cryptocurrency trading service offers competitive trading fees, fast listings, and other unique features. To continue providing its users with cost-effective services, Flipster needed to optimize the cost of its Amazon EC2 workloads.

The Flipster team also needed to overcome the hyperthreading challenges that it faced with its existing processors. At up to 80 percent, Flipster’s CPU utilization was high, and the hyperthreading feature in its x86 processors meant that the physical core in each processor functioned as if it contained multiple virtual cores.

“In my conversations with tech leaders in Singapore, they praised the price performance of Graviton,” says Jaewon An, chief technology officer of Flipster. The team decided to run a low-risk pilot to evaluate Graviton. Since Flipster uses a microservices-based architecture on Amazon Elastic Kubernetes Service (Amazon EKS)—a fully managed Kubernetes service—it ran a test by adding a few Graviton-based Amazon EC2 nodes to its Amazon EKS cluster. The promising results prompted Flipster to migrate to Graviton.

The Flipster team planned the migration in three phases: preparation, testing, and rollout. The team members attended several AWS sessions on Graviton and referred to extensive documentation to prepare for the migration. Because Graviton processors use the Arm instruction set, the team created Arm builds in its programming language repositories so that they could work with Graviton. Flipster also upgraded the continuous integration and continuous delivery pipeline to publish Arm and x86 images in one go and ran an in-house performance-benchmarking test to compare Graviton against x86. The test revealed that multithreading performance in Graviton was on par with that of x86, so the team knew that it could optimize costs without impacting performance.

During the 3-week testing phase, the team added Graviton nodes to its testing clusters and migrated the microservices one at a time. The quality assurance team did functionality and latency testing while monitoring the performance of every microservice. The test results showed that all functions worked correctly. In stress tests with high loads, the latency and throughput were noticeably more stable on Graviton than its x86 counterparts. So, the team gained the confidence to move forward.

During the rollout, the team migrated the production environment within 1 week, bringing Graviton online in waves and keeping an eye on the metrics to identify any issues. By using a dual pool of x86 and Arm images, Flipster minimized risk, because it could roll back to x86 processors if any challenges arose. In fact, the migration went smoothly, with near-zero downtime, so no rollback was required. The team did not need to recode any applications, either—and it had to tweak the paths in only 15 Docker files and its continuous integration and continuous delivery configuration—for the production environment to work on Graviton.

Flipster has migrated 95 percent of its production environment. And its critical, latency-sensitive, and input / output–heavy workloads—which include a risk engine and analytics pipelines—run smoothly on Graviton. “Our Amazon EC2 bills are down by 10 percent, and after normalizing for new features and higher traffic during the migration period, the effective cost savings are up to 20 percent,” says An.

Flipster also solved the hyperthreading challenge, because the physical cores of Graviton help simplify the calculation of capacity usage. Flipster saw a drop of up to 15 percent in CPU utilization, which meant lower saturation under peak load and higher stability. The migration also helped reduce latency by up to 40 percent—with improvements in performance—for some services. “Everything went really smoothly,” says An. “Migrating to AWS Graviton was a pleasant experience, and within 1 month we reduced both costs and latency and had more stable performance metrics.”

Flipster’s engineers have another reason to be happy about the migration. Since their development PCs also use Arm instruction sets, the build-time for new features is shorter. Before the migration, engineers had to use a virtual x86 machine to build an image of a feature and test it. That step is no longer needed because of the unified architecture.

Flipster’s users have benefited from quick feature deployment and rollout. They also have a more stable solution on Graviton, even during peak loads. Additionally, Flipster extends its cost savings to customers by maintaining zero trading fees and keeping spreads tight.

Flipster will soon migrate the remaining 5 percent of compute instances to Graviton. It plans to improve performance by setting up clusters in other AWS Regions and reduce latency by using AWS Lambda—which empowers companies to run code without thinking about servers or clusters.

Flipster also intends to explore how generative artificial intelligence (AI) can enhance its offerings. The team is working alongside AWS on the use of Amazon Bedrock—the simplest way to build and scale generative AI applications with foundation models. “AWS has been very helpful to us,” says An. “We get support quickly, and we’ve had a good experience collaborating with the AWS team.”