Real Time

New in version 13.0.0: (Mitaka)

Nova supports configuring real-time policies for instances. This builds upon the improved performance offered by CPU pinning by providing stronger guarantees for worst case scheduler latency for vCPUs.

Enabling Real-Time

Currently the creation of real-time instances is only supported when using the libvirt compute driver with a libvirt.virt_type of kvm or qemu. It requires extensive configuration of the host and this document provides but a rough overview of the changes required. Configuration will vary depending on your hardware, BIOS configuration, host and guest OS’ and application.

BIOS configuration

Configure your host BIOS as recommended in the rt-wiki page. The most important steps are:

  • Disable power management, including CPU sleep states

  • Disable SMT (hyper-threading) or any option related to logical processors

These are standard steps used in benchmarking as both sets of features can result in non-deterministic behavior.

OS configuration

This is inherently specific to the distro used, however, there are some common steps:

  • Install the real-time (preemptible) kernel (PREEMPT_RT_FULL) and real-time KVM modules

  • Configure hugepages

  • Isolate host cores to be used for instances from the kernel

  • Disable features like CPU frequency scaling (e.g. P-States on Intel processors)

RHEL and RHEL-derived distros like CentOS provide packages in their repositories to accomplish. The kernel-rt and kernel-rt-kvm packages will provide the real-time kernel and real-time KVM module, respectively, while the tuned-profiles-realtime package will provide tuned profiles to configure the host for real-time workloads. You should refer to your distro documentation for more information.

Validation

Once your BIOS and the host OS have been configured, you can validate “real-time readiness” using the hwlatdetect and rteval utilities. On RHEL and RHEL-derived hosts, you can install these using the rt-tests package. More information about the rteval tool can be found here.

Configuring a flavor or image

Changed in version 22.0.0: (Victoria)

Previously, it was necessary to specify hw:cpu_realtime_mask when hw:cpu_realtime was set to true. Starting in Victoria, it is possible to omit this when an emulator thread policy is configured using the hw:emulator_threads_policy extra spec, thus allowing all guest cores to be be allocated as real-time cores.

Changed in version 22.0.0: (Victoria)

Previously, a leading caret was necessary when specifying the value for hw:cpu_realtime_mask and omitting it would be equivalent to not setting the mask, resulting in a failure to spawn the instance.

Compared to configuring the host, configuring the guest is relatively trivial and merely requires a combination of flavor extra specs and image metadata properties, along with a suitable real-time guest OS.

Enable real-time by setting the hw:cpu_realtime flavor extra spec to true. When this is configured, it is necessary to specify where guest overhead processes should be scheduled to. This can be accomplished in one of three ways. Firstly, the hw:cpu_realtime_mask extra spec or equivalent image metadata property can be used to indicate which guest cores should be scheduled as real-time cores, leaving the remainder to be scheduled as non-real-time cores and to handle overhead processes. For example, to allocate the first two cores of an 8 core instance as the non-real-time cores:

$ openstack flavor set $FLAVOR \
    --property hw:cpu_realtime=yes \
    --property hw:cpu_realtime_mask=2-7  # so 0,1 are non-real-time

In this configuration, any non-real-time cores configured will have an implicit dedicated CPU pinning policy applied. It is possible to apply a shared policy for these non-real-time cores by specifying the mixed CPU pinning policy via the hw:cpu_policy extra spec. This can be useful to increase resource utilization of the host. For example:

$ openstack flavor set $FLAVOR \
    --property hw:cpu_policy=mixed \
    --property hw:cpu_realtime=yes \
    --property hw:cpu_realtime_mask=2-7  # so 0,1 are non-real-time and unpinned

Finally, you can explicitly offload guest overhead processes to another host core using the hw:emulator_threads_policy extra spec. For example:

$ openstack flavor set $FLAVOR \
    --property hw:cpu_realtime=yes \
    --property hw:emulator_thread_policy=share

Note

Emulator thread pinning requires additional host configuration. Refer to the documentation for more information.

In addition to configuring the instance CPUs, it is also likely that you will need to configure guest huge pages. For information on how to configure these, refer to the documentation