KubeCon + CloudNativeCon North America 2024

When you have a significant cloud footprint, you always look for performance improvements and cost reductions. So when ARM instances became commonly available on one of our providers, seemingly providing great performance at a lower cost, we had to take a closer look! In this talk, we will first describe the steps we took to make our clusters ARM-ready and a few interesting issues we encountered during our initial tests: from performance regressions due to compiler behaviors to subtle memory corruption bugs. We will then discuss new challenges, in particular how to achieve load-balancing and auto-scaling when running workloads on a mix of CPUs with different performances, and share our results. If migrating real workloads to ARM proved challenging, it was worth the effort and we now run more than 50% of our workloads on ARM.