Open Source Summit North America 2025

While containers provide isolation for CPU cycles and memory capacity, they offer limited protection against performance interference through shared CPU caches and memory bandwidth. Such contention was shown to increase application response times by 4-13x. The Linux resctrl infrastructure provides monitoring and control mechanisms, but has limitations for controlling real-world applications.

For example, child processes do not inherit their parent's resctrl groups, leaving any application that forks improperly monitored and controlled. Additionally, the current filesystem-based interface makes it difficult to build a controller that can monitor and adjust quickly enough to keep up with frequently changing application memory behavior.

This talk introduces the memory interference problem and presents new kernel mechanisms to address these limitations. A new collector enables effective control by capturing per-container measurements of cache and memory bandwidth usage at millisecond frequencies. We'll cover how the solution combines Intel RDT, AMD QoS, high-resolution timers, perf counters, and cgroups to achieve this. We'll discuss future work and opportunities for collaboration.