Nova supports configuring real-time policies for instances. This builds upon the improved performance offered by CPU pinning <cpu-topologies> by providing stronger guarantees for worst case scheduler latency for vCPUs.
Currently the creation of real-time instances is only supported when using the libvirt compute driver with a :oslo.configlibvirt.virt_type of
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.
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.
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-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.
Once your BIOS and the host OS have been configured, you can validate "real-time readiness" using the
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
Previously, it was necessary to specify :novahw:cpu_realtime_mask when realtime mode was enabled via :novahw:cpu_realtime. Starting in Victoria, it is possible to omit this when an emulator thread policy is configured using the :novahw:emulator_threads_policy extra spec, thus allowing all guest cores to be be allocated as real-time cores.
Previously, a leading caret was necessary when specifying the value for :novahw: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 :novahw:cpu_realtime flavor extra spec to
yes or a truthy value. 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 :novahw: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 <cpu-pinning-policies> applied. It is possible to apply a
shared policy for these non-real-time cores by specifying the
mixed CPU pinning policy <cpu-pinning-policies> via the :novahw: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 <emulator-thread-pinning-policies> using the :novahw:emulator_threads_policy extra spec. For example:
$ openstack flavor set $FLAVOR \ --property hw:cpu_realtime=yes \ --property hw:emulator_thread_policy=share
Emulator thread pinning requires additional host configuration. Refer to the documentation <emulator-thread-pinning-policies> 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 <huge-pages>