ae2e5650d1
Live migration is currently totally broken if a NUMA topology is present. This affects everything that's been regrettably stuffed in with NUMA topology including CPU pinning, hugepage support and emulator thread support. Side effects can range from simple unexpected performance hits (due to instances running on the same cores) to complete failures (due to instance cores or huge pages being mapped to CPUs/NUMA nodes that don't exist on the destination host). Until such a time as we resolve these issues, we should alert users to the fact that such issues exist. A workaround option is provided for operators that _really_ need the broken behavior, but it's defaulted to False to highlight the brokenness of this feature to unsuspecting operators. Change-Id: I217fba9138132b107e9d62895d699d238392e761 Signed-off-by: Stephen Finucane <sfinucan@redhat.com> Related-bug: #1289064
1.4 KiB
Advanced configuration
OpenStack clouds run on platforms that differ greatly in the capabilities that they provide. By default, the Compute service seeks to abstract the underlying hardware that it runs on, rather than exposing specifics about the underlying host platforms. This abstraction manifests itself in many ways. For example, rather than exposing the types and topologies of CPUs running on hosts, the service exposes a number of generic CPUs (virtual CPUs, or vCPUs) and allows for overcommitting of these. In a similar manner, rather than exposing the individual types of network devices available on hosts, generic software-powered network ports are provided. These features are designed to allow high resource utilization and allows the service to provide a generic cost-effective and highly scalable cloud upon which to build applications.
This abstraction is beneficial for most workloads. However, there are some workloads where determinism and per-instance performance are important, if not vital. In these cases, instances can be expected to deliver near-native performance. The Compute service provides features to improve individual instance for these kind of workloads.
pci-passthrough cpu-topologies huge-pages virtual-gpu file-backed-memory