cinder-specs/specs/kilo/filtering-weighing-with-driver-supplied-functions.rst

Filtering and weighing backends with driver supplied functions

https://blueprints.launchpad.net/cinder/+spec/filtering-weighing-with-driver-supplied-functions

Cinder currently uses a filter scheduler to filter out unqualified backends during volume operations. The default filters and weigher do a good job of filtering out the unqualified backends but there are occasions when they work less than ideally. For example, due to thin-provisioning or other advanced data manipulation mechanisms backends may have, properties such as "free capacity" will no longer be as useful when determing the best backend to use. A potential solution to this issue is to incorporate filtering and weighing (goodness) functions that can be specified for each backend as part of the scheduler decision making.

Problem description

The default filters and weighers being used by the Cinder scheduler are not determining the most qualified backends to use in some cases. If a backend is using an advanced data manipulation mechanism such as thin-provisioning the "free capacity" property, which is used by the default filters, will not be as useful or accurate when determining an ideal backend.

Another problem is that different backends use different metrics and only the manufacturer really knows what they are. This makes it hard/almost impossible to create a global filter for these metrics. An example of this would be if there was a backend array with an absolute maximum capacity of 1000 volumes. The performance of this array degrades if more than 75% of total space is actually used. There is also a maximum volume size of 500 GB. These details can be highly variable between vendors and even within different backend arrays from the same vendor.

Use Cases

Proposed change

The proposed solution for the above problem is to add two new properties to the statistics that drivers support. These two new properties will be a filter and goodness function. The details for the two new properties are described here:

* filter_function -- An equation in string form that when evaluated using
                     PyParsing will be True/False.  Used by the new filter
                     to determine if a backend should be considered by
                     the scheduler.

* goodness_function -- An equation in string form that when evaluated using
                       PyParsing will be a value between 0 and 100.  Used
                       by the new filter to which valid backend is the best
                       for the scheduler to use.

An equation can be made up of any of the following operators as long as the equation is valid:

Operations	Type
+, -, *, /, ^	standard math
not, and, or, &, |, !	logic
>, >=, <, <=, =, ==, <>, !=	equality
+, -	sign
x ? a : b	ternary
abs(x), max(x, y), min(x, y)	math helper functions

New operators could be added in the future, if desired. These ones should cover the majority of equations an admin would need.

When the scheduler attempts to find the best host it gets the stats for each backend's driver. The backend(s) with the top goodness rating will be returned to the scheduler. There is a possibility of multiple backends because there can be ties (three backends with goodness of 100). Depending on the weigher that an admin has chosen, such as capacity, random choice, etc, any potential tie can be resolved and the winning backend will be chosen.

If a backend's driver does not return both filter_function and goodness_function properties during stat reporting the new filter will default to passing that backend and give it a goodness of 0. When it comes time for the schedueler to pick the best host a 0 would convert to a "use as a last resort" rating.

If a driver chooses not to implement the new properties it will return the standard stats during reporting with no changes. Example:

{
  "volume_backend_name": "array_one",
  "pools": {
      [
        "pool_name": "my_pool",
        "percent_capacity_used": 25,
        "free_space": 5000,
        "n_vols": 50,
        ...
      ],
      ...
  },
  ...
}

Drivers that do choose to implement the new properties will need to report the two new properties back during stat reporting. An example is here:

{
  "volume_backend_name": "array_one",
  "pools": {
      [
        "pool_name": "my_pool",
        "filter_function": "(stats.n_vols < 1000) and (volume.size < 5)",
        "goodness_function": "stats.percent_capacity_used < 75 ? 100 : 25",
        "percent_capacity_used": 50,
        "free_space": 1000,
        "n_vols": 100,
        ...
      ],
      ...
  },
  ...
}

When a driver's get_volumes_stats method is called it is up to the driver to determine how to generate the equations for both the filter and goodness function properties. Some choices a driver has are to use values defined in cinder.conf, hard-code the values in the driver or not implement the properties at all. Drivers can implement other ways of achieving this, too, as long as the requirements for the filter_function and goodness_function properties are met.

Here is an example of an implementation of how using cinder.conf values would work. An admin can setup the filter and goodness function properties for each backend. Example:

[foo_1]
...
filter_function = "volume.size > 10 and volume.size <= 500"
goodness_function = "50"

[foo_2]
...
filter_function = "volume.size > 500"
goodness_function = "90"

The new filter will only focus on filtering based on the filter/weighing functions provided by a driver. If an admin desires filtering capabilities, such as for capacity, the other scheduling filters available in Cinder can be used along with this one.

In summary, implementing a new scheduling filter that allows for more control over the filtering process with a filter and goodness function will allow the best backend for a volume to be chosen correctly more often. There two potential downside to this solution. First, the equations for the filter and goodness functions are not validated until they are used by the scheduler. If there is a typo or syntax error in either equation it will not be known until the scheduler fails during evaluation. In the future some form of startup validation could be added to detect invalid equations. A simpler solution to implement for now would be to have the scheduler default to assuming an invalid filter function is passing. The goodness function would have a default of 0 incase there is an invalid equation detected in the filter or goodness functions. The second downside is that the equations that PyParsing will evaluate will require adequate documentation showing examples of how the operators work, syntax , etc. However, it should not be too hard to document this.

Alternatives

One alternative solution is to have two functions located in a file in a known location (possibly the driver itself). One function generates the filter_function result and the other generates the goodness_function result. The backend driver would still return two new stat properties, but they would now contain the import path for the needed functions. The new filter will then import the needed functions and use them during the filtering process. There could also be default base functions incase of a driver not implementing the new properties. Similar to the proposed solution, a default of a pass and goodness of 0 would be given to those backends using those drivers. Selection of backends with tieing goodness rating will be resolved in the same way as the proposed solution.

Example of custom defined filter and goodness functions:

def filter_function(stats, volume, qos_specs, extra_specs, **kwargs):
    # Return True or False
    return bool(evaluate((stats.n_vols < 1000) and (volume.size < 5)))

def goodness_function(stats, volume, qos_specs, extra_specs, **kwargs):
    # Return 0 to 100
    return clamp(evaluate(stats.percent_capacity_used < 75 ? 100 : 25))

The returned stats from a driver would now look like this:

{
  "volume_backend_name": "array_one",
  "pools": {
      [
        "pool_name": "my_pool",
        "filter_function": "path.to.my.custom.module.filter_function",
        "goodness_function": "path.to.my.custom.module.goodness_function",
        "percent_capacity_used": 50,
        "free_space": 1000,
        "n_vols": 100,
        ...
      ],
      ...
  },
  ...
}

Benefits of this solution are that the equations can be checked for typos and syntax errors more easily since it will be Python code. It would also be possible to check for these errors on startup more easily than the proposed solution. Unit tests can be developed and maintained by an adminfor the custom functions so that after an upgrade and an admin can ensure everything is working still.

Another benefit is that Python is a well documented language. The PyParsing operators used by admins to create equations will have to have documentation written to explain how to use the various operators.

A downside would be that admins would need to know how to work with python functions. Deciding how to maintain all the functions is another potential downside depending on where they are placed.

Another downside to this solution is that the functions are not coming from the driver directly. They will not be updated by or with the driver if an admin is the one creating them. Values returned by a driver can potentially change over time, e.g. a backend getting many small volumes can use the goodness value to lean towards wanting bigger volumes.

Data model impact

None

REST API impact

None

Security impact

Filtering and goodness function equations can be optionally set in cinder.conf by an admin. The equations for each function will be strings that are parsed using pyparsing. An admin can enter anything they want for the string. Pyparsing should be able to filter out and throw an exception for any dangerous strings that an admin may enter.

Notifications impact

None

Other end user impact

Admins will be able to optionally set filtering and goodness function equations in cinder.conf for backends if implemented by a driver.

Performance Impact

Depending on the complexity of a given filtering or goodness function equation there may be a slight performance decrease when it is evaluated with PyParsing. Simple equations will have minimal performance impact.

Other deployer impact

None

Developer impact

A backend's host stats, extra specs, volume info and qos specs will be available to use in the filter and goodness functions. Possibly more can be exposed in the future.

Implementation

Assignee(s)

Primary assignee:

leeantho

Other contributors:

kmartin

Work Items

• Implement evaluator code using pyparsing.
• Implement a new driver filter for scheduler that will use the evaluator.
• Update driver code to return filter and goodness function properties during stat reporting for a backend. This only needs to occur if a driver wants to implement support for this new filter.

Dependencies

The pyparsing module (2.0.1 or later).

Testing

Unit tests can be used to validate the evaluator and filter code so new Tempest tests are not needed.

Documentation Impact

• Documentation will be needed to detail what driver properties are exposed for use in the filter/goodness functions. Each driver would need to provide their own documentation for what properties they have.
• Documentation will be needed to explain the syntax for the filter and goodness function equations. Samples of using each operator would be good, too.

References

Pyparsing :: http://pyparsing.wikispaces.com/