NUMA Aware Placement and How It Can Help with Your OOMs

Learn how to solve critical out-of-memory errors in NUMA-enabled OpenStack clouds through improved resource tracking and placement strategies. Discover the fundamental flaw in current NUMA resource tracking mechanisms where individual NUMA cell usage information is stored locally in Nova rather than in Placement, leading to memory allocation issues when instances are scheduled in bulk. Explore CloudFerro's real-world experience with these problems and their implementation of the 'NUMA Topology with Resource Providers' specification as a solution. Understand how this approach, originally conceived as a performance improvement, actually serves as a necessary bug fix for high-load NUMA-enabled cloud environments, and examine the technical details of delegating work from virt driver to Placement to ensure proper resource allocation and prevent service continuity issues.