Model resources as objects¶
Adds model objects to represent the resources that may be requested for an instance.
In Nova, we have a very loose way of modeling the resources that are consumed by virtual machine instances and provided by compute nodes. The Flavor object has a number of static fields that correspond to amounts of simple resources like CPU, RAM, and local disk. We use dictionaries of key/value pairs and JSON-serialized BLOBs of data to model other types of resources, like PCIe devices or NUMA cell layouts.
The resource tracker on the compute node keeps track of the collection of resources that are consumed on the node. The ResourceTracker.old_resources attribute is a dictionary with a bunch of random nested dictionaries. Some of these nested dictionaries include the ‘stats’ dict for the extensible resource tracker, various ‘pci_devices’, ‘pci_stats’, and ‘pci_passthrough_devices’ things, a ‘numa_topology’ blob that stores a JSON-serialized representation of an object in nova.virt.hardware, and a ‘metrics’ dictionary with completely unstructured and undocumented key/value pairs. In addition to these unstructured nested dictionaries, the ResourceTracker.old_resources dictionary contains top-level keys including some that match the simple resource types that a Flavor object exposes:
local_gb_used: Amount of disk in GB used on the compute node
local_gb: Total GB of local disk capacity the compute node provides
free_disk_gb: Calculated amount of disk the compute node has available
vcpus_used: Number of vCPUs consumed on the compute node
vcpus: Total number of vCPUs the compute node provides
free_vcpus: Calculated number of vCPUs the compute node has available
memory_mb_used: Amount of RAM in MB used on the compute node
memory_mb: Total MB of RAM capacity the compute node provides
free_ram_mb: Calculated amount of RAM the compute node has available
running_vms: Number of virtual machine instances running on the node
current_workload: Some calculated value of the workload on the node
Unfortunately, none of the above is documented in the code, and in order to add new features to the scheduler, people just continue to add these free-form keys and nested dictionaries to the dictionary. This makes communicating actual usage amounts to the scheduler error-prone: the resource tracker calls the scheduler_client.update_resource_stats() method, passing in this unstructured, unversioned dictionary of information as-is. This means the scheduler interface is incredibly fragile since the interface can be altered on a whim by any developer who decides to add a new key to the free-form dictionary of resources. Typos in resource dictionary keys can be very easy to miss in code reviews, and frankly, there is virtually no functional testing for a lot of the edge case code in the resource tracker around the extensible resource tracker.
In addition to the problem of fragile interfaces, the free-form nature of the resources dictionary has meant that different resources are tracked in different ways. PCI resources are tracked one way, NUMA topology usage is tracked in a different way, CPU/RAM/disk is tracked differently, and any resources modeled in the complete free-for-all of the extensible resource tracker are tracked in an entirely different way, using plugins that modify a supplied ‘stats’ nested dictionary.
An example of the mess this has created in the resource tracker can be seen here:
def _update(self, context, values): """Update partial stats locally and populate them to Scheduler.""" self._write_ext_resources(values) # NOTE(pmurray): the stats field is stored as a json string. The # json conversion will be done automatically by the ComputeNode object # so this can be removed when using ComputeNode. values['stats'] = jsonutils.dumps(values['stats']) if not self._resource_change(values): return if "service" in self.compute_node: del self.compute_node['service'] # NOTE(sbauza): Now the DB update is asynchronous, we need to locally # update the values self.compute_node.update(values) # Persist the stats to the Scheduler self._update_resource_stats(context, values) if self.pci_tracker: self.pci_tracker.save(context)
If resources were actually modeled consistently, the above code would look like this instead:
def _update(self, context, resources): if not self._resource_change(resources): return # Notify the scheduler about changed resources scheduler_client.update_usage_for_compute_node( context, self.compute_node.id, resources)
Similarly, the following code in the existing resource tracker:
def _update_usage(self, context, resources, usage, sign=1): mem_usage = usage['memory_mb'] overhead = self.driver.estimate_instance_overhead(usage) mem_usage += overhead['memory_mb'] resources['memory_mb_used'] += sign * mem_usage resources['local_gb_used'] += sign * usage.get('root_gb', 0) resources['local_gb_used'] += sign * usage.get('ephemeral_gb', 0) # free ram and disk may be negative, depending on policy: resources['free_ram_mb'] = (resources['memory_mb'] - resources['memory_mb_used']) resources['free_disk_gb'] = (resources['local_gb'] - resources['local_gb_used']) resources['running_vms'] = self.stats.num_instances self.ext_resources_handler.update_from_instance(usage, sign) # Calculate the numa usage free = sign == -1 updated_numa_topology = hardware.get_host_numa_usage_from_instance( resources, usage, free) resources['numa_topology'] = updated_numa_topology
would instead look like this:
def _update_usage(self, context, resources, usage, sign=1): for resource_type, amount in usage.items(): resources[resource_type].consume(amount*sign)
Nova contributors wish to extend the functionality of the scheduler and intend to break the scheduler out into the Gantt project. In order to do this effectively, the internal interfaces around the resource tracker and the scheduler must be cleaned up to use structured objects.
This blueprint is part of the scheduler refactoring effort defined as a priority for the Kilo release.
Modeling requested and used resource amounts is the foundational building block that must be done first before any further refactoring or cleanup of the scheduler or resource tracker interfaces.
This blueprint encompasses the addition of a set of model objects:
A set of classes that represent amounts of different types of resources
A set of classes that represent usage of different types of resources
A set of classes that represent different types of resources
These amount, usage, and resource type classes will be nova.objects object classes, and will enable Nova evolve, in a versioned manner, the way that it tracks resources and exposes usage data about consumed resources.
The goals of the extensible resource tracker was to put in place a framework that allowed adding new resource types and allowed accounting for those resources in different ways. While this blueprint does indeed remove the extensible resource tracker, because these amount, usage, and resource type classes are being added as nova.object objects, we will gain the ability to evolve and enhance our resource tracking with object versioning and adding new resource type classes. This approach will provide the flexibility that the ERT intended, but with the stability that the nova objects system.
The resource tracker code will then be converted to use the above classes when representing usage of all resources on a compute node.
A new conductor RPC API method update_compute_node_resource_usages method will be added that gets a nova.objects.ComputeNode object and a nova.objects.ResourceUsageCollection object and updates the compute node database record with the resource usage amounts/data coming from the resource tracker.
Note that we do NOT propose any changes at this time to the database schema of the compute_nodes table or the fields in nova.objects.ComputeNode, however we do add translation methods to nova.objects.ComputeNode that will be able to construct a set of UsageSpec objects for each resource type on the compute node as well as set the existing fields on the nova.objects.ComputeNode object from a supplied set of UsageSpec objects.
Data model impact¶
None. The objects added in this blueprint are not stored in a database. These objects are a replacement for an unstructured nested dictionary that is currently used to represent resource amounts.
REST API impact¶
Other end user impact¶
Other deployer impact¶
The extensible resource tracker will be removed when this blueprint is completed.
Once this blueprint is completed, code handling the construction of the request_spec will be much more structured, and much of the spaghetti code in the resource tracker around the ERT, PCI tracker, and NUMA topology quirks will go away.
There will be a generalized usage spec class interface that looks like this:
class UsageSpec(object): """Represents the used amount of a particular type of resource.""" def update(self, amount_spec): """Update the used amount of resources by the supplied amount. :param amount_spec: `AmountSpec` to modify the usage with. """ raise NotImplementedError def has_room_for(self, amount_spec): """Determine if there is room to fit the supplied amount of resources. :param amount_spec: `AmountSpec` to determine if there is room for. :returns True if the supplied requested amount of resources is able to be consumed on the node, False otherwise. If the supplied `amount_spec` is negative, returns False . """ raise NotImplementedError
Note that each concrete specialization of the UsageSpec class must be able to handle overcommit ratios for the type of resource that it handles.
With the idea that all requested resources for an instance should be able to be compared to all resource usage records for a compute node in the same way, using code that looks like this:
for resource_type, amount in request_spec.resources.items(): if compute_node.usages[resource_type].has_room_for(amount): # do something... perhaps claim resources on the compute # node, which might eventually call: compute_node.usages[resource_type].update(amount)
The conductor RPC API should have a new update_compute_node_resource_usages() method added to it that passes the new nova.objects.ResourceUsageCollection object to the conductor, that the conductor decomposes and updates the compute node records in the database.
- Primary assignee:
Add classes for requested and usage amount representation
Add classes for resource type representation
Add a class representing a collection of requested resource amounts
Add a class representing a collection of resource usages
Add translation methods to nova.objects.ComputeNode to construct and decompose a set of UsageSpec objects w/ unit tests
Convert resource tracker to use usage amount objects instead of triples of free/total/used amounts in key/value pairs in a dictionary for the non-PCI, non-ERT, non-NUMA resources.
Remove the extensible resource tracker code.
Convert resource tracker to use usage amount objects instead of ‘numa_topology’ key and nova.virt.hardware.VirtNUMATopology object in the old_resources dictionary.
Convert resource tracker to use usage amount objects instead of ‘pci_devices’ and ‘pci_passthrough_devices’ keys and a nova.pci.pci_stats.PciDeviceStats object in the pci_tracker attribute of the resource tracker.
Convert the virt driver’s get_available_resources method to return a dictionary of resource objects.
Add a new conductor RPC API (and scheduler client) method called update_compute_node_resource_usages() that supplies the new nova.objects.ResourceUsageCollection object representation of resource usages.
Deprecate the old update_resource_stats() conductor RPC API method.
Convert the scheduler’s HostStateManager to utilize the new ComputeNode.get_resource_usages() method and ComputeNode.update_from_resource_usages method.
Add developer reference documentation for how resources are modeled.
New unit tests for the objects will be added. The existing unit tests of resource tracker will be overhauled in the patch set that converts the resource tracker to use the new resource object models instead of its current free-form dictionary of things.
There are currently no developer reference docs that explain how the different resources are tracked within Nova. Developer reference material that explains the new resource type and amount classes will be delivered as a part of this blueprint.
This blueprint is part of an overall effort to clean up, version, and stabilize the interfaces between the nova-api, nova-scheduler, nova-conductor and nova-compute daemons that involve scheduling and resource decisions.
resource-objects <– this blueprint