Getting On The Nested Magic Train 1

https://storyboard.openstack.org/#!/story/2005575

This spec describes a cluster of Placement API work to support several interrelated use cases for Train around:

  • Modeling complex trees such as NUMA layouts, multiple devices, networks.

  • Requesting affinity [1] between/among the various providers/allocations in allocation candidates against such layouts.

  • Describing granular groups more richly to facilitate the above.

  • Requesting candidates based on traits that are not necessarily associated with resources.

An additional spec, for a feature known as can_split has been separated out to its own spec to ensure that any delay in it does not impact these features, which are less controversial.

Principles

In developing this design, some fundamental concepts have come to light. These are not really changes from the existing architecture, but understanding them becomes more important in light of the changes introduced herein.

Resource versus Provider Traits

The database model associates traits with resource providers, not with inventories of resource classes. However, conceptually there are two different categories of traits to consider.

Resource Traits are tied to specific resources. For example, HW_CPU_X86_AVX2 describes a characteristic of VCPU (or PCPU) resources.

Provider Traits are characteristics of a provider, regardless of the resources it provides. For example, COMPUTE_VOLUME_MULTI_ATTACH is a capability of a compute host, not of any specific resource inventory. HW_NUMA_ROOT describes NUMA affinity among all the resources in the inventories of that provider and all its descendants. CUSTOM_PHYSNET_PUBLIC indicates connectivity to the public network, regardless of whether the associated resources are VF, PF, VNIC, etc.; and regardless of whether those resources reside on the provider marked with the trait or on its descendants.

This distinction becomes important when deciding how to model. Resource traits need to “follow” their resource class. For example, HW_CPU_X86_AVX2 should be on the provider of VCPU (or PCPU) resource, whether that’s the root or a NUMA child. On the other hand, provider traits must stick to their provider, regardless of where resources inventories are placed. For example, COMPUTE_VOLUME_MULTI_ATTACH should always be on the root provider, as the root provider conceptually represents “the compute host”.

Alternative: “Traits Flow Down”: There have been discussions around a provider implicitly inheriting the traits of its parent (and therefore all its ancestors). This would (mostly) allow us not to think about the distinction between “resource” and “provider” traits. We ultimately decided against this by a hair, mainly because of this:

It makes no sense to say my PGPU is capable of MULTI_ATTACH

In addition, IIUC, there are SmartNICs [1] that have CPUs on cards. If someone will want to report/model those CPUs in placement, they will be scared that CPU traits on compute side flow down to those CPUs on NIC despite they are totally different CPUs.

[1] https://www.netronome.com/products/smartnic/overview/

…and because we were able to come up with other satisfactory solutions to our use cases.

Group-Specific versus Request-Wide Query Parameters

granular resource requests introduced a divide between GET /allocation_candidates query parameters which apply to a particular request group

  • resources[$S]

  • required[$S]

  • member_of[$S]

  • in_tree[$S]

…and those which apply to the request as a whole

  • limit

  • group_policy

This has been fairly obvious thus far; but this spec introduces concepts (such as root_required and same_subtree) that make it important to keep this distinction in mind. Moving forward, we should consider whether new features and syntax additions make more sense to be group-specific or request-wide.

Proposed change

All changes are to the GET /allocation_candidates operation via new microversions, one per feature described below.

arbitrary group suffixes

Use case: Client code managing request groups for different kinds of resources - which will often come from different providers - may reside in different places in the codebase. For example, the management of compute resources vs. networks vs. accelerators. However, there still needs to be a way for the consuming code to express relationships (such as affinity) among these request groups. For this purpose, API consumers wish to be able to use conventions for request group identifiers. It would also be nice for development and debugging purposes if these designations had some element of human readability.

(Merged) code is here: https://review.opendev.org/#/c/657419/

Granular groups are currently restricted to using integer suffixes. We will change this so they can be case-sensitive strings up to 64 characters long comprising alphanumeric (either case), underscore, and hyphen.

For purposes of documentation (and this spec), we’ll rename the “unnumbered” group to “unspecified” or “unsuffixed”, and anywhere we reference “numbered” groups we can call them “suffixed” or “granular” (I think this label is already used in some places).

same_subtree

Use case: I want to express affinity between/among allocations in separate request groups. For example, that a VGPU come from a GPU affined to the NUMA node that provides my VCPU and MEMORY_MB; or that multiple network VFs come from the same NIC.

A new same_subtree query parameter will be accepted. The value is a comma-separated list of request group suffix strings $S. Each must exactly match a suffix on a granular group somewhere else in the request. Importantly, the identified request groups need not have a resources$S (see resourceless request groups).

We define “same subtree” as “all of the resource providers satisfying the request group must be rooted at one of the resource providers satisfying the request group”. Or put another way: “one of the resource providers satisfying the request group must be the direct ancestor of all the other resource providers satisfying the request group”.

For example, given a model like:

            +--------------+
            | compute node |
            +-------+------+
                    |
          +---------+----------+
          |                    |
+---------+--------+ +---------+--------+
| numa0            | | numa1            |
| VCPU: 4 (2 used) | | VCPU: 4          |
| MEMORY_MB: 2048  | | MEMORY_MB: 2048  |
+---+--------------+ +---+----------+---+
    |                    |          |
+---+----+           +---+---+  +---+---+
|fpga0_0 |           |fpga1_0|  |fpga1_1|
|FPGA:1  |           |FPGA:1 |  |FPGA:1 |
+--------+           +-------+  +-------+

to request “two VCPUs, 512MB of memory, and one FPGA from the same NUMA node,” my request could include:

?resources_COMPUTE=VCPU:2,MEMORY_MB:512
&resources_ACCEL=FPGA:1
# NOTE: The suffixes include the leading underscore!
&same_subtree=_COMPUTE,_ACCEL

This will produce candidates including:

- numa0: {VCPU:2, MEMORY_MB:512}, fpga0_0: {FPGA:1}
- numa1: {VCPU:2, MEMORY_MB:512}, fpga1_0: {FPGA:1}
- numa1: {VCPU:2, MEMORY_MB:512}, fpga1_1: {FPGA:1}

but not:

- numa0: {VCPU:2, MEMORY_MB:512}, fpga1_0: {FPGA:1}
- numa0: {VCPU:2, MEMORY_MB:512}, fpga1_1: {FPGA:1}
- numa1: {VCPU:2, MEMORY_MB:512}, fpga0_0: {FPGA:1}

The same_subtree query parameter is request-wide, but may be repeated. Each grouping is treated independently.

Anti-affinity

There were discussions about supporting ! syntax in same_subtree to express anti-affinity (e.g. same_subtree=$X,!$Y meaning “resources from group $Y shall not come from the same subtree as resources from group $X”). This shall be deferred to a future release.

resourceless request groups

Use case: When making use of same_subtree, I want to be able to identify a provider as a placeholder in the subtree structure even if I don’t need any resources from that provider.

It is currently a requirement that a resources$S exist for all $S in a request. This restriction shall be removed such that a request group may exist e.g. with only required$S or member_of$S.

There must be at least one resources or resources$S somewhere in the request, otherwise there will be no inventory to allocate and thus no allocation candidates. If neither is present a 400 response will be returned.

Furthermore, resourceless request groups must be used with same_subtree. That is, the suffix for each resourceless request group must feature in a same_subtree somewhere in the request. Otherwise a 400 response will be returned. (The reasoning for this restriction is explained below.)

For example, given a model like:

            +--------------+
            | compute node |
            +-------+------+
                    |
        +-----------+-----------+
        |                       |
  +-----+-----+           +-----+-----+
  |nic1       |           |nic2       |
  |HW_NIC_ROOT|           |HW_NIC_ROOT|
  +-----+-----+           +-----+-----+
        |                       |
   +----+----+            +-----+---+
   |         |            |         |
+--+--+   +--+--+      +--+--+   +--+--+
|pf1_1|   |pf1_2|      |pf2_1|   |pf2_2|
|NET1 |   |NET2 |      |NET1 |   |NET2 |
|VF:4 |   |VF:4 |      |VF:2 |   |VF:2 |
+-----+   +-----+      +-----+   +-----+

a request such as the following, meaning, “Two VFs from the same NIC, one on each of network NET1 and NET2,” is legal:

?resources_VIF_NET1=VF:1
&required_VIF_NET1=NET1
&resources_VIF_NET2=VF:1
&required_VIF_NET2=NET2
# NOTE: there is no resources_NIC_AFFINITY
&required_NIC_AFFINITY=HW_NIC_ROOT
&same_subtree=_VIF_NET1,_VIF_NET2,_NIC_AFFINITY

The returned candidates will include:

- pf1_1: {VF:1}, pf1_2: {VF:1}
- pf2_1: {VF:1}, pf2_2: {VF:1}

but not:

- pf1_1: {VF:1}, pf2_2: {VF:1}
- pf2_1: {VF:1}, pf1_2: {VF:1}

Why enforce resourceless + same_subtree?

Taken by itself (without same_subtree), a resourceless request group intuitively means, “There must exist in the solution space a resource provider that satisfies these constraints.” But what does “solution space” mean? Clearly it’s not the same as solution path, or we wouldn’t be able to use it to add resourceless providers to that solution path. So it must encompass at least the entire non-sharing tree around the solution path. Does it also encompass sharing providers associated via aggregate? What would that mean?

Since we have not identified any real use cases for resourceless without same_subtree (other than root_member_of – see below) making this an error allows us to not have to deal with these questions.

root_required

Use case: I want to limit allocation candidates to trees whose root provider has (or does not have) certain traits. For example, I want to limit candidates to only multi-attach-capable hosts; or preserve my Windows-licensed hosts for special use.

A new root_required query parameter will be accepted. The value syntax is identical to that of required[$S]: that is, it accepts a comma-delimited list of trait names, each optionally prefixed with ! to indicate “forbidden” rather than “required”.

This is a request-wide query parameter designed for provider traits specifically on the root provider of the non-sharing tree involved in the allocation candidate. That is, regardless of any group-specific constraints, and regardless of whether the root actually provides resource to the request, results will be filtered such that the root of the non-sharing tree conforms to the constraints specified in root_required.

root_required may not be repeated.

The fact that this feature is (somewhat awkwardly) restricted to “…trees whose root provider …” deserves some explanation. This is to fill a gap in use cases that cannot be adequately covered by other query parameters.

  • To land on a tree (host) with a given trait anywhere in its hierarchy, resourceless request groups without same_subtree could be used. However, there is no way to express the “forbidden” side of this in a way that makes sense:

    • A resourceless required$S=!FOO would simply ensure that a provider anywhere in the tree does not have FOO - which would end up not being restrictive as intended in most cases.

    • We could define “resourceless forbidden” to mean “nowhere in the tree”, but this would be inconsistent and hard to explain.

  • To ensure that the desired trait is present (or absent) in the result set, it would be necessary to attach the trait to a group whose resource constraints will be satisfied by the provider possessing (or lacking) that trait.

    • This requires the API consumer to understand too much about how the provider trees are modeled; and

    • It doesn’t work in heterogeneous environments where such provider traits may or may not stick with providers of a specific resource class.

    This could possibly be mitigated by careful use of same_subtree, but that again requires deep understanding of the tree model, and also confuses the meaning of same_subtree and resource versus provider traits.

  • The traits flow down concept described earlier could help here; but that would still entail attaching provider traits to a particular request group. Which one? Because the trait isn’t associated with a specific resource, it would be arbitrary and thus difficult to explain and justify.

Alternative: “Solution Path”: A more general solution was discussed whereby we would define a “solution path” as: The set of resource providers which satisfy all the request groups *plus* all the ancestors of those providers, up to the root. This would allow us to introduce a request-wide query parameter such as solution_path_required. The idea would be the same as root_required, but the specified trait constraints would be applied to all providers in the “solution path” (required traits must be present somewhere in the solution path; forbidden traits must not be present anywhere in the solution path).

This alternative was rejected because:

  • Describing the “solution path” concept to API consumers would be hard.

  • We decided the only real use cases where the trait constraints needed to be applied to providers other than the root could be satisfied (and more naturally) in other ways.

This section was the result of long discussions in IRC and on the review for this spec

root_member_of

Note

When this spec was initially written it was not clear whether there was immediate need to implement this feature. This turned out to be the case. The feature was not implemented in the Train cycle. It will be revisted in the future if needed.

Use case: I want to limit allocation candidates to trees whose root provider is (or is not) a member of a certain aggregate. For example, I want to limit candidates to only hosts in (or not in) a specific availability zone.

Note

We “need” this because of the restriction that resourceless request groups must be used with same_subtree. Without that restriction, a resourceless member_of would match a provider anywhere in the tree, including the root.

root_member_of is conceptually identical to root_required, but for aggregates. Like member_of[$S], root_member_of supports in:, and can be repeated (in contrast to [root_]required[$S]).

Default group_policy to none

A single isolate setting that applies to the whole request has consistently been shown to be inadequate/confusing/frustrating for all but the simplest anti-affinity use cases. We’re not going to get rid of group_policy, but we’re going to make it no longer required, defaulting to none. This will allow us to get rid of at least one hack in nova and provide a clearer user experience, while still allowing us to satisfy simple NUMA use cases. In the future a granular isolation syntax should make it possible to satisfy more complex scenarios.

(Future) Granular Isolation

Note

This is currently out of scope, but we wanted to get it written down.

The features elsewhere in this spec allow us to specify affinity pretty richly. But anti-affinity (within a provider tree - not between providers) is still all (group_policy=isolate) or nothing (group_policy=none). We would like to be able to express anti-affinity between/among subsets of the suffixed groups in the request.

We propose a new request-wide query parameter key isolate. The value is a comma-separated list of request group suffix strings $S. Each must exactly match a suffix on a granular group somewhere else in the request. This works on resourceless request groups as well as those with resources. It is mutually exclusive with the group_policy query parameter: 400 if both are specified.

The effect is the resource providers satisfying each group $S must satisfy only their respective group $S.

At one point I thought it made sense for isolate to be repeatable. But now I can’t convince myself that isolate={set1}&isolate={set2} can ever produce an effect different from isolate={set1|set2}. Perhaps it’s because different isolates could be coming from different parts of the calling code?

Another alternative would be to isolate the groups from each other but not from other groups, in which case repeating isolate could be meaningful. But confusing. Thought will be needed.

Interactions

Some discussion on these can be found in the neighborhood of http://eavesdrop.openstack.org/irclogs/%23openstack-placement/%23openstack-placement.2019-05-10.log.html#t2019-05-10T22:02:43

group_policy + same_subtree

group_policy=isolate forces the request groups identified in same_subtree to be satisfied by different providers, whereas group_policy=none would also allow same_subtree to degenerate to “same provider”.

For example, given the following model:

            +--------------+
            | compute node |
            +-------+------+
                    |
        +-----------+-----------+
        |                       |
  +-----+-----+           +-----+-----+
  |nic1       |           |nic2       |
  |HW_NIC_ROOT|           |HW_NIC_ROOT|
  +-----+-----+           +-----+-----+
        |                       |
   +----+----+                 ...
   |         |
+--+--+   +--+--+
|pf1_1|   |pf1_2|
|VF:4 |   |VF:4 |
+-----+   +-----+

a request for “Two VFs from different PFs on the same NIC”:

?resources_VIF1=VF:1
&resources_VIF2=VF:1
&required_NIC_AFFINITY=HW_NIC_ROOT
&same_subtree=_VIF1,_VIF2,_NIC_AFFINITY
&group_policy=isolate

will return only one candidate:

- pf1_1: {VF:1}, pf1_2: {VF:1}

whereas the same request with group_policy=none, meaning “Two VFs from the same NIC”:

?resources_VIF1=VF:1
&resources_VIF2=VF:1
&required_NIC_AFFINITY=HW_NIC_ROOT
&same_subtree=_VIF1,_VIF2,_NIC_AFFINITY
&group_policy=none

will return two additional candidates where both VFs are satisfied by the same provider:

- pf1_1: {VF:1}, pf1_2: {VF:1}
- pf1_1: {VF:2}
- pf1_2: {VF:2}

group_policy + resourceless request groups

Resourceless request groups are treated the same as any other for the purposes of group_policy:

  • If your resourceless request group is suffixed, group_policy=isolate means the provider satisfying the resourceless request group will not be able to satisfy any other suffixed group.

  • If your resourceless request group is unsuffixed, it can be satisfied by any provider in the tree, since the unsuffixed group isn’t isolated (even with group_policy=isolate). This is important because there are cases where we want to require certain traits (usually provider traits), and don’t want to figure out which other request group might be requesting resources from the same provider.

same_subtree + resourceless request groups

These must be used together – see Why enforce resourceless + same_subtree?

Impacts

Data model impact

There should be no changes to database table definitions, but the implementation will almost certainly involve adding/changing database queries.

There will also likely be changes to python-side objects representing meta-objects used to manage information between the database and the REST layer. However, the data models for the JSON payloads in the REST layer itself will be unaffected.

Performance Impact

The work for same_subtree will probably (at least initially) be done on the python side as additional filtering under _merge_candidates. This could have some performance impact especially on large data sets. Again, we should optimize requests without same_subtree, where same_subtree refers to only one group, where no nested providers exist in the database, etc.

Resourceless request groups may add a small additional burden to database queries, but it should be negligible. It should be relatively rare in the wild for a resourceless request group to be satisfied by a provider that actually provides no resource to the request, though there are cases where a resourceless request group would be useful even though the provider does provide resources to the request.

Documentation Impact

The new query parameters will be documented in the API reference.

Microversion paperwork will be done.

Modeling with Provider Trees will be updated (and/or split off of).

Security impact

None

Other end user impact

None

Other deployer impact

None

Developer impact

None

Upgrade impact

None

Implementation

Assignee(s)

  • cdent

  • tetsuro

  • efried

  • others

Dependencies

None

Testing

Code for a gabbi fixture with some complex and interesting characteristics is merged here: https://review.opendev.org/#/c/657463/

Lots of functional testing, primarily via gabbi, will be included.

It wouldn’t be insane to write some PoC consuming code on the nova side to validate assumptions and use cases.

References

…are inline

History

Revisions

Release Name

Description

Train

Introduced