Filter Scheduler

Filter Scheduler

The Filter Scheduler supports filtering and weighting to make informed decisions on where a new instance should be created. This Scheduler supports working with Compute Nodes only.

Filtering

../_images/filtering-workflow-1.png

During its work Filter Scheduler iterates over all found compute nodes, evaluating each against a set of filters. The list of resulting hosts is ordered by weighers. The Scheduler then chooses hosts for the requested number of instances, choosing the most weighted hosts. For a specific filter to succeed for a specific host, the filter matches the user request against the state of the host plus some extra magic as defined by each filter (described in more detail below).

If the Scheduler cannot find candidates for the next instance, it means that there are no appropriate hosts where that instance can be scheduled.

The Filter Scheduler has to be quite flexible to support the required variety of filtering and weighting strategies. If this flexibility is insufficient you can implement your own filtering algorithm.

There are many standard filter classes which may be used (nova.scheduler.filters):

  • AllHostsFilter - does no filtering. It passes all the available hosts.

  • ImagePropertiesFilter - filters hosts based on properties defined on the instance’s image. It passes hosts that can support the properties specified on the image used by the instance.

  • AvailabilityZoneFilter - filters hosts by availability zone. It passes hosts matching the availability zone specified in the instance properties. Use a comma to specify multiple zones. The filter will then ensure it matches any zone specified.

  • ComputeCapabilitiesFilter - checks that the capabilities provided by the host compute service satisfy any extra specifications associated with the instance type. It passes hosts that can create the specified instance type.

    If an extra specs key contains a colon (:), anything before the colon is treated as a namespace and anything after the colon is treated as the key to be matched. If a namespace is present and is not capabilities, the filter ignores the namespace. For example capabilities:cpu_info:features is a valid scope format. For backward compatibility, when a key doesn’t contain a colon (:), the key’s contents are important. If this key is an attribute of HostState object, like free_disk_mb, the filter also treats the extra specs key as the key to be matched. If not, the filter will ignore the key.

    The extra specifications can have an operator at the beginning of the value string of a key/value pair. If there is no operator specified, then a default operator of s== is used. Valid operators are:

    * = (equal to or greater than as a number; same as vcpus case)
    * == (equal to as a number)
    * != (not equal to as a number)
    * >= (greater than or equal to as a number)
    * <= (less than or equal to as a number)
    * s== (equal to as a string)
    * s!= (not equal to as a string)
    * s>= (greater than or equal to as a string)
    * s> (greater than as a string)
    * s<= (less than or equal to as a string)
    * s< (less than as a string)
    * <in> (substring)
    * <all-in> (all elements contained in collection)
    * <or> (find one of these)
    
    Examples are: ">= 5", "s== 2.1.0", "<in> gcc", "<all-in> aes mmx", and "<or> fpu <or> gpu"
    

    some of attributes that can be used as useful key and their values contains:

    * free_ram_mb (compared with a number, values like ">= 4096")
    * free_disk_mb (compared with a number, values like ">= 10240")
    * host (compared with a string, values like: "<in> compute","s== compute_01")
    * hypervisor_type (compared with a string, values like: "s== QEMU", "s== powervm")
    * hypervisor_version (compared with a number, values like : ">= 1005003", "== 2000000")
    * num_instances (compared with a number, values like: "<= 10")
    * num_io_ops (compared with a number, values like: "<= 5")
    * vcpus_total (compared with a number, values like: "= 48", ">=24")
    * vcpus_used (compared with a number, values like: "= 0", "<= 10")
    
  • AggregateInstanceExtraSpecsFilter - checks that the aggregate metadata satisfies any extra specifications associated with the instance type (that have no scope or are scoped with aggregate_instance_extra_specs). It passes hosts that can create the specified instance type. The extra specifications can have the same operators as ComputeCapabilitiesFilter. To specify multiple values for the same key use a comma. E.g., “value1,value2”. All hosts are passed if no extra_specs are specified.

  • ComputeFilter - passes all hosts that are operational and enabled.

  • CoreFilter - DEPRECATED; filters based on CPU core utilization. It passes hosts with sufficient number of CPU cores.

  • AggregateCoreFilter - filters hosts by CPU core number with per-aggregate cpu_allocation_ratio setting. If no per-aggregate value is found, it will fall back to the global default cpu_allocation_ratio. If more than one value is found for a host (meaning the host is in two different aggregates with different ratio settings), the minimum value will be used.

  • IsolatedHostsFilter - filter based on filter_scheduler.isolated_images, filter_scheduler.isolated_hosts and filter_scheduler.restrict_isolated_hosts_to_isolated_images flags.

  • JsonFilter - allows simple JSON-based grammar for selecting hosts.

  • RamFilter - DEPRECATED; filters hosts by their RAM. Only hosts with sufficient RAM to host the instance are passed.

  • AggregateRamFilter - filters hosts by RAM with per-aggregate ram_allocation_ratio setting. If no per-aggregate value is found, it will fall back to the global default ram_allocation_ratio. If more than one value is found for a host (meaning the host is in two different aggregates with different ratio settings), the minimum value will be used.

  • DiskFilter - DEPRECATED; filters hosts by their disk allocation. Only hosts with sufficient disk space to host the instance are passed. disk_allocation_ratio setting. The virtual disk to physical disk allocation ratio, 1.0 by default. The total allowed allocated disk size will be physical disk multiplied this ratio.

  • AggregateDiskFilter - filters hosts by disk allocation with per-aggregate disk_allocation_ratio setting. If no per-aggregate value is found, it will fall back to the global default disk_allocation_ratio. If more than one value is found for a host (meaning the host is in two or more different aggregates with different ratio settings), the minimum value will be used.

  • NumInstancesFilter - filters compute nodes by number of running instances. Nodes with too many instances will be filtered. filter_scheduler.max_instances_per_host setting. Maximum number of instances allowed to run on this host. The host will be ignored by the scheduler if more than filter_scheduler.max_instances_per_host already exists on the host.

  • AggregateNumInstancesFilter - filters hosts by number of instances with per-aggregate filter_scheduler.max_instances_per_host setting. If no per-aggregate value is found, it will fall back to the global default filter_scheduler.max_instances_per_host. If more than one value is found for a host (meaning the host is in two or more different aggregates with different max instances per host settings), the minimum value will be used.

  • IoOpsFilter - filters hosts by concurrent I/O operations on it. hosts with too many concurrent I/O operations will be filtered. filter_scheduler.max_io_ops_per_host setting. Maximum number of I/O intensive instances allowed to run on this host, the host will be ignored by scheduler if more than filter_scheduler.max_io_ops_per_host instances such as build/resize/snapshot etc are running on it.

  • AggregateIoOpsFilter - filters hosts by I/O operations with per-aggregate filter_scheduler.max_io_ops_per_host setting. If no per-aggregate value is found, it will fall back to the global default :oslo.config:option:`filter_scheduler.max_io_ops_per_host. If more than one value is found for a host (meaning the host is in two or more different aggregates with different max io operations settings), the minimum value will be used.

  • PciPassthroughFilter - Filter that schedules instances on a host if the host has devices to meet the device requests in the ‘extra_specs’ for the flavor.

  • SimpleCIDRAffinityFilter - allows a new instance on a host within the same IP block.

  • DifferentHostFilter - allows the instance on a different host from a set of instances.

  • SameHostFilter - puts the instance on the same host as another instance in a set of instances.

  • RetryFilter - filters hosts that have been attempted for scheduling. Only passes hosts that have not been previously attempted.

  • AggregateTypeAffinityFilter - limits instance_type by aggregate.

    This filter passes hosts if no instance_type key is set or the instance_type aggregate metadata value contains the name of the instance_type requested. The value of the instance_type metadata entry is a string that may contain either a single instance_type name or a comma separated list of instance_type names. e.g. ‘m1.nano’ or “m1.nano,m1.small”

  • ServerGroupAntiAffinityFilter - This filter implements anti-affinity for a server group. First you must create a server group with a policy of ‘anti-affinity’ via the server groups API. Then, when you boot a new server, provide a scheduler hint of ‘group=<uuid>’ where <uuid> is the UUID of the server group you created. This will result in the server getting added to the group. When the server gets scheduled, anti-affinity will be enforced among all servers in that group.

  • ServerGroupAffinityFilter - This filter works the same way as ServerGroupAntiAffinityFilter. The difference is that when you create the server group, you should specify a policy of ‘affinity’.

  • AggregateMultiTenancyIsolation - isolate tenants in specific aggregates. To specify multiple tenants use a comma. Eg. “tenant1,tenant2”

  • AggregateImagePropertiesIsolation - isolates hosts based on image properties and aggregate metadata. Use a comma to specify multiple values for the same property. The filter will then ensure at least one value matches.

  • MetricsFilter - filters hosts based on metrics weight_setting. Only hosts with the available metrics are passed.

  • NUMATopologyFilter - filters hosts based on the NUMA topology requested by the instance, if any.

Now we can focus on these standard filter classes in some detail. We’ll skip the simplest ones, such as AllHostsFilter, CoreFilter and RamFilter, because their functionality is relatively simple and can be understood from the code. For example class RamFilter has the next realization:

class RamFilter(filters.BaseHostFilter):
    """Ram Filter with over subscription flag"""

    def host_passes(self, host_state, filter_properties):
        """Only return hosts with sufficient available RAM."""
        instance_type = filter_properties.get('instance_type')
        requested_ram = instance_type['memory_mb']
        free_ram_mb = host_state.free_ram_mb
        total_usable_ram_mb = host_state.total_usable_ram_mb
        used_ram_mb = total_usable_ram_mb - free_ram_mb
        return total_usable_ram_mb * FLAGS.ram_allocation_ratio  - used_ram_mb >= requested_ram

Here ram_allocation_ratio means the virtual RAM to physical RAM allocation ratio (it is 1.5 by default).

The AvailabilityZoneFilter looks at the availability zone of compute node and availability zone from the properties of the request. Each compute service has its own availability zone. So deployment engineers have an option to run scheduler with availability zones support and can configure availability zones on each compute host. This class’s method host_passes returns True if availability zone mentioned in request is the same on the current compute host.

The ImagePropertiesFilter filters hosts based on the architecture, hypervisor type and virtual machine mode specified in the instance. For example, an instance might require a host that supports the ARM architecture on a qemu compute host. The ImagePropertiesFilter will only pass hosts that can satisfy this request. These instance properties are populated from properties defined on the instance’s image. E.g. an image can be decorated with these properties using glance image-update img-uuid --property architecture=arm --property hypervisor_type=qemu Only hosts that satisfy these requirements will pass the ImagePropertiesFilter.

ComputeCapabilitiesFilter checks if the host satisfies any extra_specs specified on the instance type. The extra_specs can contain key/value pairs. The key for the filter is either non-scope format (i.e. no : contained), or scope format in capabilities scope (i.e. capabilities:xxx:yyy). One example of capabilities scope is capabilities:cpu_info:features, which will match host’s cpu features capabilities. The ComputeCapabilitiesFilter will only pass hosts whose capabilities satisfy the requested specifications. All hosts are passed if no extra_specs are specified.

ComputeFilter is quite simple and passes any host whose compute service is enabled and operational.

Now we are going to IsolatedHostsFilter. There can be some special hosts reserved for specific images. These hosts are called isolated. So the images to run on the isolated hosts are also called isolated. The filter checks if filter_scheduler.isolated_images flag named in instance specifications is the same as the host specified in filter_scheduler.isolated_hosts. Isolated hosts can run non-isolated images if the flag filter_scheduler.restrict_isolated_hosts_to_isolated_images is set to false.

DifferentHostFilter - method host_passes returns True if the host to place an instance on is different from all the hosts used by a set of instances.

SameHostFilter does the opposite to what DifferentHostFilter does. host_passes returns True if the host we want to place an instance on is one of the hosts used by a set of instances.

SimpleCIDRAffinityFilter looks at the subnet mask and investigates if the network address of the current host is in the same sub network as it was defined in the request.

JsonFilter - this filter provides the opportunity to write complicated queries for the hosts capabilities filtering, based on simple JSON-like syntax. There can be used the following operations for the host states properties: =, <, >, in, <=, >=, that can be combined with the following logical operations: not, or, and. For example, the following query can be found in tests:

['and',
    ['>=', '$free_ram_mb', 1024],
    ['>=', '$free_disk_mb', 200 * 1024]
]

This query will filter all hosts with free RAM greater or equal than 1024 MB and at the same time with free disk space greater or equal than 200 GB.

Many filters use data from scheduler_hints, that is defined in the moment of creation of the new server for the user. The only exception for this rule is JsonFilter, that takes data from the schedulers HostState data structure directly. Variable naming, such as the $free_ram_mb example above, should be based on those attributes.

The RetryFilter filters hosts that have already been attempted for scheduling. It only passes hosts that have not been previously attempted. If a compute node is raising an exception when spawning an instance, then the compute manager will reschedule it by adding the failing host to a retry dictionary so that the RetryFilter will not accept it as a possible destination. That means that if all of your compute nodes are failing, then the RetryFilter will return 0 hosts and the scheduler will raise a NoValidHost exception even if the problem is related to 1:N compute nodes. If you see that case in the scheduler logs, then your problem is most likely related to a compute problem and you should check the compute logs.

The NUMATopologyFilter considers the NUMA topology that was specified for the instance through the use of flavor extra_specs in combination with the image properties, as described in detail in the related nova-spec document:

and try to match it with the topology exposed by the host, accounting for the ram_allocation_ratio and cpu_allocation_ratio for over-subscription. The filtering is done in the following manner:

  • Filter will attempt to pack instance cells onto host cells.
  • It will consider the standard over-subscription limits for each host NUMA cell, and provide limits to the compute host accordingly (as mentioned above).
  • If instance has no topology defined, it will be considered for any host.
  • If instance has a topology defined, it will be considered only for NUMA capable hosts.

Configuring Filters

To use filters you specify two settings:

The default values for these settings in nova.conf are:

--filter_scheduler.available_filters=nova.scheduler.filters.all_filters
--filter_scheduler.enabled_filters=ComputeFilter,AvailabilityZoneFilter,ComputeCapabilitiesFilter,ImagePropertiesFilter,ServerGroupAntiAffinityFilter,ServerGroupAffinityFilter

With this configuration, all filters in nova.scheduler.filters would be available, and by default the ComputeFilter, AvailabilityZoneFilter, ComputeCapabilitiesFilter, ImagePropertiesFilter, ServerGroupAntiAffinityFilter, and ServerGroupAffinityFilter would be used.

Each filter selects hosts in a different way and has different costs. The order of filter_scheduler.enabled_filters affects scheduling performance. The general suggestion is to filter out invalid hosts as soon as possible to avoid unnecessary costs. We can sort filter_scheduler.enabled_filters items by their costs in reverse order. For example, ComputeFilter is better before any resource calculating filters like RamFilter, CoreFilter.

In medium/large environments having AvailabilityZoneFilter before any capability or resource calculating filters can be useful.

Writing Your Own Filter

To create your own filter you must inherit from BaseHostFilter and implement one method: host_passes. This method should return True if a host passes the filter and return False elsewhere. It takes two parameters (named arbitrarily as host_state and spec_obj):

  • the HostState object allows to get attributes of the host.
  • the RequestSpec object describes the user request, including the flavor, the image and the scheduler hints.

For further details about each of those objects and their corresponding attributes, please refer to the codebase (at least by looking at the other filters code) or ask for help in the #openstack-nova IRC channel.

As an example, nova.conf could contain the following scheduler-related settings:

--scheduler.driver=nova.scheduler.FilterScheduler
--filter_scheduler.available_filters=nova.scheduler.filters.all_filters
--filter_scheduler.available_filters=myfilter.MyFilter
--filter_scheduler.enabled_filters=RamFilter,ComputeFilter,MyFilter

Note

When writing your own filter, be sure to add it to the list of available filters and enable it in the default filters. The “all_filters” setting only includes the filters shipped with nova.

With these settings, nova will use the FilterScheduler for the scheduler driver. All of the standard nova filters and MyFilter are available to the FilterScheduler, but just the RamFilter, ComputeFilter, and MyFilter will be used on each request.

Weights

Filter Scheduler uses the so-called weights during its work. A weigher is a way to select the best suitable host from a group of valid hosts by giving weights to all the hosts in the list.

In order to prioritize one weigher against another, all the weighers have to define a multiplier that will be applied before computing the weight for a node. All the weights are normalized beforehand so that the multiplier can be applied easily. Therefore the final weight for the object will be:

weight = w1_multiplier * norm(w1) + w2_multiplier * norm(w2) + ...

A weigher should be a subclass of weights.BaseHostWeigher and they can implement both the weight_multiplier and _weight_object methods or just implement the weight_objects method. weight_objects method is overridden only if you need access to all objects in order to calculate weights, and it just return a list of weights, and not modify the weight of the object directly, since final weights are normalized and computed by weight.BaseWeightHandler.

The Filter Scheduler weighs hosts based on the config option filter_scheduler.weight_classes, this defaults to nova.scheduler.weights.all_weighers, which selects the following weighers:

  • RAMWeigher Compute weight based on available RAM on the compute node. Sort with the largest weight winning. If the multiplier, filter_scheduler.ram_weight_multiplier, is negative, the host with least RAM available will win (useful for stacking hosts, instead of spreading). Starting with the Stein release, if per-aggregate value with the key filter_scheduler.ram_weight_multiplier is found, this value would be chosen as the ram weight multiplier. Otherwise, it will fall back to the filter_scheduler.ram_weight_multiplier. If more than one value is found for a host in aggregate metadata, the minimum value will be used.

  • CPUWeigher Compute weight based on available vCPUs on the compute node. Sort with the largest weight winning. If the multiplier, filter_scheduler.cpu_weight_multiplier, is negative, the host with least CPUs available will win (useful for stacking hosts, instead of spreading). Starting with the Stein release, if per-aggregate value with the key filter_scheduler.cpu_weight_multiplier is found, this value would be chosen as the cpu weight multiplier. Otherwise, it will fall back to the filter_scheduler.cpu_weight_multiplier. If more than one value is found for a host in aggregate metadata, the minimum value will be used.

  • DiskWeigher Hosts are weighted and sorted by free disk space with the largest weight winning. If the multiplier is negative, the host with less disk space available will win (useful for stacking hosts, instead of spreading). Starting with the Stein release, if per-aggregate value with the key filter_scheduler.disk_weight_multiplier is found, this value would be chosen as the disk weight multiplier. Otherwise, it will fall back to the filter_scheduler.disk_weight_multiplier. If more than one value is found for a host in aggregate metadata, the minimum value will be used.

  • MetricsWeigher This weigher can compute the weight based on the compute node host’s various metrics. The to-be weighed metrics and their weighing ratio are specified in the configuration file as the followings:

    metrics_weight_setting = name1=1.0, name2=-1.0
    

    Starting with the Stein release, if per-aggregate value with the key metrics_weight_multiplier is found, this value would be chosen as the metrics weight multiplier. Otherwise, it will fall back to the metrics.weight_multiplier. If more than one value is found for a host in aggregate metadata, the minimum value will be used.

  • IoOpsWeigher The weigher can compute the weight based on the compute node host’s workload. The default is to preferably choose light workload compute hosts. If the multiplier is positive, the weigher prefer choosing heavy workload compute hosts, the weighing has the opposite effect of the default. Starting with the Stein release, if per-aggregate value with the key filter_scheduler.io_ops_weight_multiplier is found, this value would be chosen as the IO ops weight multiplier. Otherwise, it will fall back to the filter_scheduler.io_ops_weight_multiplier. If more than one value is found for a host in aggregate metadata, the minimum value will be used.

  • PCIWeigher Compute a weighting based on the number of PCI devices on the host and the number of PCI devices requested by the instance. For example, given three hosts - one with a single PCI device, one with many PCI devices, and one with no PCI devices - nova should prioritise these differently based on the demands of the instance. If the instance requests a single PCI device, then the first of the hosts should be preferred. Similarly, if the instance requests multiple PCI devices, then the second of these hosts would be preferred. Finally, if the instance does not request a PCI device, then the last of these hosts should be preferred.

    For this to be of any value, at least one of the PciPassthroughFilter or NUMATopologyFilter filters must be enabled.

    Configuration Option:
     [filter_scheduler] pci_weight_multiplier. Only positive values are allowed for the multiplier as a negative value would force non-PCI instances away from non-PCI hosts, thus, causing future scheduling issues.

    Starting with the Stein release, if per-aggregate value with the key filter_scheduler.pci_weight_multiplier is found, this value would be chosen as the pci weight multiplier. Otherwise, it will fall back to the filter_scheduler.pci_weight_multiplier. If more than one value is found for a host in aggregate metadata, the minimum value will be used.

  • ServerGroupSoftAffinityWeigher The weigher can compute the weight based on the number of instances that run on the same server group. The largest weight defines the preferred host for the new instance. For the multiplier only a positive value is allowed for the calculation. Starting with the Stein release, if per-aggregate value with the key filter_scheduler.soft_affinity_weight_multiplier is found, this value would be chosen as the soft affinity weight multiplier. Otherwise, it will fall back to the filter_scheduler.soft_affinity_weight_multiplier. If more than one value is found for a host in aggregate metadata, the minimum value will be used.

  • ServerGroupSoftAntiAffinityWeigher The weigher can compute the weight based on the number of instances that run on the same server group as a negative value. The largest weight defines the preferred host for the new instance. For the multiplier only a positive value is allowed for the calculation. Starting with the Stein release, if per-aggregate value with the key filter_scheduler.soft_anti_affinity_weight_multiplier is found, this value would be chosen as the soft anti-affinity weight multiplier. Otherwise, it will fall back to the filter_scheduler.soft_anti_affinity_weight_multiplier. If more than one value is found for a host in aggregate metadata, the minimum value will be used.

  • BuildFailureWeigher Weigh hosts by the number of recent failed boot attempts. It considers the build failure counter and can negatively weigh hosts with recent failures. This avoids taking computes fully out of rotation. Starting with the Stein release, if per-aggregate value with the key filter_scheduler.build_failure_weight_multiplier is found, this value would be chosen as the build failure weight multiplier. Otherwise, it will fall back to the filter_scheduler.build_failure_weight_multiplier. If more than one value is found for a host in aggregate metadata, the minimum value will be used.

Filter Scheduler makes a local list of acceptable hosts by repeated filtering and weighing. Each time it chooses a host, it virtually consumes resources on it, so subsequent selections can adjust accordingly. It is useful if the customer asks for a large block of instances, because weight is computed for each instance requested.

../_images/filtering-workflow-2.png

At the end Filter Scheduler sorts selected hosts by their weight and attempts to provision instances on the chosen hosts.

P.S.: you can find more examples of using Filter Scheduler and standard filters in nova.tests.scheduler.

Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.