18 KiB

Raw Blame History

Distributed Scheduler

The Scheduler is akin to a Dating Service. Requests for the creation of new instances come in and the most applicable Compute nodes are selected from a large pool of potential candidates. In a small deployment we may be happy with the currently available Chance Scheduler which randomly selects a Host from the available pool. Or if you need something a little more fancy you may want to use the Availability Zone Scheduler, which selects Compute hosts from a logical partitioning of available hosts (within a single Zone).

But for larger deployments a more complex scheduling algorithm is required. Additionally, if you are using Zones in your Nova setup, you'll need a scheduler that understand how to pass instance requests from Zone to Zone.

This is the purpose of the Distributed Scheduler (DS). The DS utilizes the Capabilities of a Zone and its component services to make informed decisions on where a new instance should be created. When making this decision it consults not only all the Compute nodes in the current Zone, but the Compute nodes in each Child Zone. This continues recursively until the ideal host is found.

So, how does this all work?

This document will explain the strategy employed by the BaseScheduler, which is the base for all schedulers designed to work across zones, and its derivations. You should read the devguide/zones documentation before reading this.

Costs & Weights

When deciding where to place an Instance, we compare a Weighted Cost for each Host. The Weighting, currently, is just the sum of each Cost. Costs are nothing more than integers from 0 - max_int. Costs are computed by looking at the various Capabilities of the Host relative to the specs of the Instance being asked for. Trying to put a plain vanilla instance on a high performance host should have a very high cost. But putting a vanilla instance on a vanilla Host should have a low cost.

Some Costs are more esoteric. Consider a rule that says we should prefer Hosts that don't already have an instance on it that is owned by the user requesting it (to mitigate against machine failures). Here we have to look at all the other Instances on the host to compute our cost.

An example of some other costs might include selecting:

a GPU-based host over a standard CPU
a host with fast ethernet over a 10mbps line
a host that can run Windows instances
a host in the EU vs North America
etc

This Weight is computed for each Instance requested. If the customer asked for 1000 instances, the consumed resources on each Host are "virtually" depleted so the Cost can change accordingly.

nova.scheduler.base_scheduler.BaseScheduler

As we explained in the Zones documentation, each Scheduler has a ZoneManager object that collects "Capabilities" about child Zones and each of the services running in the current Zone. The BaseScheduler uses this information to make its decisions.

Here is how it works:

The compute nodes are filtered and the nodes remaining are weighed.

Filtering the hosts is a simple matter of ensuring the compute node has ample resources (CPU, RAM, Disk, etc) to fulfil the request.

Weighing of the remaining compute nodes assigns a number based on their suitability for the request.

The same request is sent to each child Zone and step #1 is done there too. The resulting weighted list is returned to the parent.

The parent Zone sorts and aggregates all the weights and a final build plan is constructed.

The build plan is executed upon. Concurrently, instance create requests are sent to each of the selected hosts, be they local or in a child zone. Child Zones may forward the requests to their child Zones as needed.

BaseScheduler by itself is not capable of handling all the provisioning itself. You should also specify the filter classes and weighting classes to be used in determining which host is selected for new instance creation.

Filtering and Weighing

The filtering (excluding compute nodes incapable of fulfilling the request) and weighing (computing the relative "fitness" of a compute node to fulfill the request) rules used are very subjective operations ... Service Providers will probably have a very different set of filtering and weighing rules than private cloud administrators. The filtering and weighing aspects of the BaseScheduler are flexible and extensible.

Requesting a new instance

Prior to the BaseScheduler, to request a new instance, a call was made to nova.compute.api.create(). The type of instance created depended on the value of the InstanceType record being passed in. The InstanceType determined the amount of disk, CPU, RAM and network required for the instance. Administrators can add new InstanceType records to suit their needs. For more complicated instance requests we need to go beyond the default fields in the InstanceType table.

nova.compute.api.create() performed the following actions:

it validated all the fields passed into it.
it created an entry in the Instance table for each instance requested
it put one run_instance message in the scheduler queue for each instance requested
the schedulers picked off the messages and decided which compute node should handle the request.
the run_instance message was forwarded to the compute node for processing and the instance is created.
it returned a list of dicts representing each of the Instance records (even if the instance has not been activated yet). At least the instance_ids are valid.

Generally, the simplest schedulers (like ChanceScheduler and AvailabilityZoneScheduler) only operate in the current Zone. They have no concept of child Zones.

The problem with this approach is each request is scattered amongst each of the schedulers. If we are asking for 1000 instances, each scheduler gets the requests one-at-a-time. There is no possability of optimizing the requests to take into account all 1000 instances as a group. We call this Single-Shot vs. All-at-Once.

For the BaseScheduler we need to use the All-at-Once approach. We need to consider all the hosts across all the Zones before deciding where they should reside. In order to handle this we have a new method nova.compute.api.create_all_at_once(). This method does things a little differently:

it validates all the fields passed into it.
it creates a single reservation_id for all of instances created. This is a UUID.
it creates a single run_instance request in the scheduler queue
a scheduler picks the message off the queue and works on it.
the scheduler sends off an OS API POST /zones/select command to each child Zone. The BODY payload of the call contains the request_spec.
the child Zones use the request_spec to compute a weighted list for each instance requested. No attempt to actually create an instance is done at this point. We're only estimating the suitability of the Zones.
if the child Zone has its own child Zones, the /zones/select call will be sent down to them as well.
Finally, when all the estimates have bubbled back to the Zone that initiated the call, all the results are merged, sorted and processed.
Now the instances can be created. The initiating Zone either forwards the run_instance message to the local Compute node to do the work, or it issues a POST /servers call to the relevant child Zone. The parameters to the child Zone call are the same as what was passed in by the user.
The reservation_id is passed back to the caller. Later we explain how the user can check on the status of the command with this reservation_id.

The Catch

This all seems pretty straightforward but, like most things, there's a catch. Zones are expected to operate in complete isolation from each other. Each Zone has its own AMQP service, database and set of Nova services. But for security reasons Zones should never leak information about the architectural layout internally. That means Zones cannot leak information about hostnames or service IP addresses outside of its world.

When POST /zones/select is called to estimate which compute node to use, time passes until the POST /servers call is issued. If we only passed the weight back from the select we would have to re-compute the appropriate compute node for the create command ... and we could end up with a different host. Somehow we need to remember the results of our computations and pass them outside of the Zone. Now, we could store this information in the local database and return a reference to it, but remember that the vast majority of weights are going to be ignored. Storing them in the database would result in a flood of disk access and then we have to clean up all these entries periodically. Recall that there are going to be many, many select calls issued to child Zones asking for estimates.

Instead, we take a rather innovative approach to the problem. We encrypt all the child Zone internal details and pass them back the to parent Zone. In the case of a nested Zone layout, each nesting layer will encrypt the data from all of its children and pass that to its parent Zone. In the case of nested child Zones, each Zone re-encrypts the weighted list results and passes those values to the parent. Every Zone interface adds another layer of encryption, using its unique key.

Once a host is selected, it will either be local to the Zone that received the initial API call, or one of its child Zones. In the latter case, the parent Zone it simply passes the encrypted data for the selected host back to each of its child Zones during the POST /servers call as an extra parameter. If the child Zone can decrypt the data, then it is the correct Zone for the selected host; all other Zones will not be able to decrypt the data and will discard the request. This is why it is critical that each Zone has a unique value specified in its config in `--build_plan_encryption_key`: it controls the ability to locate the selected host without having to hard-code path information or other identifying information. The child Zone can then act on the decrypted data and either go directly to the Compute node previously selected if it is located in that Zone, or repeat the process with its child Zones until the target Zone containing the selected host is reached.

Throughout the nova.api.openstack.servers, nova.api.openstack.zones, nova.compute.api.create* and nova.scheduler.base_scheduler code you'll see references to blob and child_blob. These are the encrypted hints about which Compute node to use.

Reservation IDs

The OpenStack API allows a user to list all the instances they own via the GET /servers/ command or the details on a particular instance via GET /servers/###. This mechanism is usually sufficient since OS API only allows for creating one instance at a time, unlike the EC2 API which allows you to specify a quantity of instances to be created.

NOTE: currently the GET /servers command is not Zone-aware since all operations done in child Zones are done via a single administrative account. Therefore, asking a child Zone to GET /servers would return all the active instances ... and that would not be what the user intended. Later, when the Keystone Auth system is integrated with Nova, this functionality will be enabled.

We could use the OS API 1.1 Extensions mechanism to accept a num_instances parameter, but this would result in a different return code. Instead of getting back an Instance record, we would be getting back a reservation_id. So, instead, we've implemented a new command POST /zones/boot command which is nearly identical to POST /servers except that it takes a num_instances parameter and returns a reservation_id. Perhaps in OS API 2.x we can unify these approaches.

Finally, we need to give the user a way to get information on each of the instances created under this reservation_id. Fortunately, this is still possible with the existing GET /servers command, so long as we add a new optional reservation_id parameter.

python-novaclient will be extended to support both of these changes.

Host Filter

As we mentioned earlier, filtering hosts is a very deployment-specific process. Service Providers may have a different set of criteria for filtering Compute nodes than a University. To faciliate this the nova.scheduler.filters module supports a variety of filtering strategies as well as an easy means for plugging in your own algorithms.

The filter used is determined by the --default_host_filters flag, which points to a Python Class. By default this flag is set to [AllHostsFilter] which simply returns all available hosts. But there are others:

InstanceTypeFilter provides host filtering based on the memory and disk size specified in the InstanceType record passed into run_instance.

JSONFilter filters hosts based on simple JSON expression grammar. Using a LISP-like JSON structure the caller can request instances based on criteria well beyond what InstanceType specifies. See nova.tests.test_host_filter for examples.

To create your own HostFilter the user simply has to derive from nova.scheduler.filters.AbstractHostFilter and implement two methods: instance_type_to_filter and filter_hosts. Since Nova is currently dependent on the InstanceType structure, the instance_type_to_filter method should take an InstanceType and turn it into an internal data structure usable by your filter. This is for backward compatibility with existing OpenStack and EC2 API calls. If you decide to create your own call for creating instances not based on Flavors or InstanceTypes you can ignore this method. The real work is done in filter_hosts which must return a list of host tuples for each appropriate host. The set of available hosts is in the host_list parameter passed into the call as well as the filter query. The host tuple contains (<hostname>, <additional data>) where <additional data> is whatever you want it to be. By default, it is the capabilities reported by the host.

Cost Scheduler Weighing

Every BaseScheduler subclass should also override the weigh_hosts method. This takes the list of filtered hosts (generated by the filter_hosts method) and returns a list of weight dicts. The weight dicts must contain two keys: weight and hostname where weight is simply an integer (lower is better) and hostname is the name of the host. The list does not need to be sorted, this will be done by the BaseScheduler when all the results have been assembled.

Simple Scheduling Across Zones ----------------------------The BaseScheduler uses the default filter_hosts method, which will use either any filters specified in the request's filter parameter, or, if that is not specified, the filters specified in the FLAGS.default_host_filters setting. Its weight_hosts method simply returns a weight of 1 for all hosts. But, from this, you can see calls being routed from Zone to Zone and follow the flow of things.

The --scheduler_driver flag is how you specify the scheduler class name.

Flags

All this Zone and Distributed Scheduler stuff can seem a little daunting to configure, but it's actually not too bad. Here are some of the main flags you should set in your nova.conf file:

--allow_admin_api=true
--enable_zone_routing=true
--zone_name=zone1
--build_plan_encryption_key=c286696d887c9aa0611bbb3e2025a45b
--scheduler_driver=nova.scheduler.base_scheduler.BaseScheduler
--default_host_filter=nova.scheduler.filters.AllHostsFilter

--allow_admin_api must be set for OS API to enable the new /zones/* commands. --enable_zone_routing must be set for OS API commands such as create(), pause() and delete() to get routed from Zone to Zone when looking for instances. --zone_name is only required in child Zones. The default Zone name is nova, but you may want to name your child Zones something useful. Duplicate Zone names are not an issue. build_plan_encryption_key is the SHA-256 key for encrypting/decrypting the Host information when it leaves a Zone. Be sure to change this key for each Zone you create. Do not duplicate keys. scheduler_driver is the real workhorse of the operation. For Distributed Scheduler, you need to specify a class derived from nova.scheduler.base_scheduler.BaseScheduler. default_host_filter is the host filter to be used for filtering candidate Compute nodes.

Some optional flags which are handy for debugging are:

--connection_type=fake
--verbose

Using the Fake virtualization driver is handy when you're setting this stuff up so you're not dealing with a million possible issues at once. When things seem to working correctly, switch back to whatever hypervisor your deployment uses.

18 KiB Raw Blame History