cri-resource-manager
cri-resource-manager copied to clipboard
intel
Kubernetes Container Runtime Interface proxy service with hardware resource aware workload placement policies
**Describe the solution you'd like** We have most of the cpufreq controls implemented for CPU classes. We have two left to allow to be controlled by cluster admin: energy_performance_preference and...
In #816 a multi-memtier variant of memtier benchmarking was added. This issue is to track and co-ordinate the follow-up improvements. Two points were made in the #816, carrying those over...
Make sure that e2e tests can be run with NRI support enabled in cri-rm and containerd. As of this writing, support with cri-o is not tested. This PR is still...
**Describe the bug** Looks like the RDT example configurations (in `docs/` and `sample-configs/`) are outdated. They follow the pre-goresctrl spec which is slightly different from what is expected. **Expected behavior**...
**Describe the bug** cri-resmgr crashes when balloons policy is activated with `AvailableResources.CPU: 0`. **Expected behavior** Print an error **To Reproduce** ``` $ cat > conf.yaml
## Release Process - [ ] In the issue description, add a changelog section, describing changes since the last release. - Local release preparations - [ ] Perform mandatory internal...
This PR improves startup-time resource discovery of running containers. It does this primarily by sending each discovered container a `ContainerStatus` request with the `Verbose` flag set, digging out the bits...
**Describe the bug** cri-resmgr fails to properly adjust resources in container creation/update requests, if the node is using cgroup v2. **Expected behavior** cri-resmgr autodetects if cgroup v2 is in use...
- Taking triplet coverage into account improves data coverage without increasing the number of parameter sets. - As a downside, this implementation is naive and may cause performance problems with...
An idea to be investigated further: currently the lowest pool level is a NUMA node. Could we get further benefit with certain workloads if we would split the lowest level...