v0.7.0
This release of NRI Reference Plugins brings in new features and important bug fixes.
Highlights
- Topology-aware and balloons resource policies now support soft-disabling of hyperthreads per container. This improves the performance of some classes of workloads. Both policies support new pod annotation:
and the balloons policy has new balloon-type option
hide-hyperthreads.resource-policy.nri.io/container.<CONTAINER-NAME>: "true"
hideHyperthreads
that soft-disables hyperthreads on all containers assigned to a balloon of this type. - The topology-aware policy supports pinning containers to high-bandwidth memory (HBM), or both HBM and DRAM, when pods are annotated with
memory-type.resource-policy.nri.io/container.<CONTAINER-NAME>: hbm memory-type.resource-policy.nri.io/container.<CONTAINER-NAME>: hbm,dram
- Automatic hardware topology hint generation has been fixed in the topology-aware policy. For instance, if a container uses a PCI device, the policy prefers pinning the container to CPUs and memory that are close to the device.
What's New
- balloons: hideHyperthreads balloon type option and annotation by @askervin in #338
- topology-aware: add support for hide-hyperthreads annotation. by @askervin in #331
What Changed
- topology-aware: don't ignore HBM memory nodes without close CPUs. by @klihub in #329
- topology-aware: relax NUMA node topology checks. by @klihub in #336
- resmgr: exit when ttrpc connection goes down. by @klihub in #319
- cpuallocator: don't filter based on single CoreKind. by @klihub in #345
- sysfs,cpuallocator: fix CPU cluster discovery. by @klihub in #337
- sysfs: survive NUMA nodes without memory. by @klihub in #339
- sysfs: allow non-uniform thread count. by @klihub in #340
- helm: flip podPriorityClassNodeCritical to true. by @klihub in #312
- config-manager: allow configuring NRI timeouts. by @klihub in #318
New Contributors
Full Changelog: v0.5.0...v0.7.0