Competition for memory bandwidth and CPU caches between containers can increase application response times by 5x to 14x, even with CPU and memory limits in place. It can be triggered by common events like garbage collection, and existing observability tools do not collect the metrics to detect it. As it manifests as latency SLO violations, operators often scale out and run at low utilization: expensive, and only marginally improves response times.

CPU performance counters can detect memory interference. However since interference events are frequent and short-lived, detecting them requires high-frequency measurements, which is challenging due to jitter and overhead.

This session first presents the causes of memory noisy neighbors, real-world patterns that trigger it, and the benefits of mitigation. We then show how a new open source collector combines CPU performance counters, eBPF, and high-resolution timers to identify noisy neighbors in Kubernetes.

In cloud-native environments, application performance often degrades due to contention over shared resources such as CPU caches and memory bandwidth. Current container technologies lack mechanisms to isolate these resources, which compels operators to maintain low utilization by scaling out their deployments.

This session explores strategies used by hyperscalers like Google and Alibaba Cloud to mitigate such performance interference. We will review their published methodologies, extracting key principles that could guide the development of a Kubernetes-native performance isolator. Participants will gain insights into the design trade-offs and operational impacts of these tools. Additionally, we will discuss integration strategies for deploying such isolators in existing Kubernetes environments, aiming to optimize resource utilization while preserving application performance.