wiki:GAPI_AM_API_DRAFT/Adopted

Version 1 (modified by Aaron Helsinger, 10 years ago) (diff)

--

GENI Aggregate Manager API Adopted Revisions

This page documents adopted revisions to the GENI Aggregate Manager API, but not yet incorporated in a new version of the AM API.

The current officially adopted version of the API is 3 and is documented on the main API page.

Current proposals for the next version of the AM API are on the AM API Draft wiki page.

Proposals on this page are generally targeted at AM API version 4, but may be implemented in earlier versions of the AM API as appropriate.

Change Set C: Update()

This change was adopted at the GEC15 coding sprint for inclusion in API v4. Aggregates are permitted to include this method in the API v3 implementations.

Add an ability for experimenters to modify their allocated resources at an aggregate without deleting (and possibly losing) existing resource allocations.

This change was briefly discussed at GEC13, discussed at the GEC14 coding sprint, and the GEC15 AM Topics and Coding Sprint.

Motivation

A common complaint among experimenters about the current AM API is that there is no way to atomically add/remove and modify resources at an aggregate. AM API v3 allows aggregates to support adding or deleting some resources from the slice at the aggregate, but not all aggregates can support that without the ability to also modify other resources. For example, you cannot add a link without adding an interface to one or more nodes, and there is no way to modify the node without first deleting it. This proposal aims to address that, by introducing a method to update the slice at an aggregate.

The SFA calls for an UpdateSlice method, "to request that additional resources—as specified in the RSpec—be allocated to the slice".

In the PlanetLab implementation of the SFA, UpdateSliver is in fact a synonym for CreateSliver - the server will ensure that your allocated resources match your request RSpec, adding, removing and modifying resources as needed. It immediately allocates and boots nodes to match the request RSpec.

The ProtoGENI CMV2 API has UpdateSliver, which is described as the way to "Request a change of resuorces for an existing sliver. The new set of resources that are desired are specified in the rspec." At ProtoGENI as at PlanetLab, this method takes the full RSpec description of resources the experimenter wants, and the server computes the difference with what the experimenter already has. At ProtoGENI though, this method returns a ticket. The experimenter must then redeem the ticket to actually acquire the resources.

This topic was discussed at the GEC12 AM API session and on the GENI dev mailing list (in October and November).

This topic was also discussed on the geni-dev mailing list here: http://lists.geni.net/pipermail/dev/2012-March/000588.html and here: http://lists.geni.net/pipermail/dev/2012-March/000643.html and elsewhere in March 2012.

A competing related proposal from Gary Wong is here: http://www.protogeni.net/trac/protogeni/wiki/AM_API_proposals

Jon Duerig made a proposal prior to GEC14, which was discussed at the coding sprint. This proposal is a revised version of that proposal.

Proposal Summary

This proposal adds an Update() method, a Cancel() method, and a new geni_allocation_state: geni_updating. The proposal includes a new state guarantee RSpec extension, a new geni_cancelled option to Describe(), and calls for some minor changes to GENI v3 RSpecs.

Update()

Begins a transaction to modify resources which are currently in the geni_allocated or geni_provisioned state. Update() itself changes only the internal allocation of slivers, not their operational state. It is therefore fast and synchronous.

struct Update(string urns[], struct credentials[], struct geni.rspec rspec, struct options)

This call accepts the geni_best_effort option; when supplied, the aggregate should attempt to partially satisfy the request.

This call accepts the geni_end_time option; when supplied, the aggregate may attempt to renew the slivers being created or modified to the requested time, according to aggregate-local policy.

This change adds a new geni_updating state, for slivers that were geni_provisioned but have been updated. Here is a state diagram:

Since Update() changes the allocation state only, it has an immediate effect on slivers currently in the geni_allocated or geni_updating states. However, updating slivers in the geni_provisioned state is a two-step process involving first changing the allocation via Update() and then committing those changes with Provision().

The input RSpec is a complete request specification for all slivers included in the urns list, as you would like the slivers to exist when the call is fully committed. Any difference in the update RSpec from the current slivers shall be interpreted as a modification request, including removing interfaces or links, or deleting entire slivers.

The RSpec argument to Update() should be a request RSpec, but RSpecs passing the manifest schema are allowed. Resources in the RSpec should include the sliver_id tag when the experimenter is requesting a modification to the given resource. Resources without a sliver_id are to be interpreted as new resource requests.

If a sliver is defined in the RSpec and contained in the urns list, (either by URN or by indirect inclusion because the slice URN is in the urns list), then the user is requesting that the sliver be modified. If the sliver was in the geni_allocated state, then the allocation of that sliver is changed immediately and the sliver remains in a geni_allocated state. If the sliver was in the geni_provisioned state, then the operational state of that sliver is preserved and it is placed in the geni_updating state where it can be Provision()ed to implement the modification or Cancel()led.

If the RSpec does not define one or more slivers which are specified in the urns list (that is, does not include resources containing the given sliver_id tag), then the user is requesting that the given sliver(s) should be deleted. If the sliver was in the geni_allocated state, the allocated sliver is deleted and the sliver ceases to exist (becomes geni_unallocated). If the sliver was in the geni_provisioned state, then the operational state of that sliver is preserved and it is placed in the geni_updating allocation state where it can be Provision()ed to delete the sliver or Cancel()led to preserve it.

Any slivers which are not defined in the RSpec and do not appear in the urns list will remain unchanged. They remain in their current allocation state.

Any resources requested in the RSpec that are not specified in the urns list will be instantiated and be in the geni_allocated state, if the aggregate has the resources available and otherwise can provide the resources following the aggregate's policy for Allocate.

Note that at least one urn must be provided in the urns argument (slice or sliver). If a sliver urn is supplied and that sliver is unknown or expired, then Update shall result in an error (e.g. SEARCHFAILED, EXPIRED or ERROR geni_code) (unless geni_best_effort is true, in which case the method may succeed, but return a geni_error for each sliver that failed). If Update is called with a slice urn which identifies a slice with no existing reservation at this aggregate, then Update behaves like Allocate. Not however that if the supplied RSpec identifies any specific slivers (via inclusion of a sliver_id), then those sections are treated as new resource requests.

After a successful Update()

After a successful call to Update(), all slivers in the urns list which began in the geni_provisioned state will be in the geni_updating state (and their configuration will revert in the case of a Cancel() call described below). This includes those slivers which are marked for deletion as described above. These slivers will not actually be deleted until the Provision() call.

Slivers which were geni_provisioned and are now geni_updating will retain their prior existing expiration time. If the expiration time is reached before the sliver is Provision()ed, then the resources are freed and the sliver is geni_unallocated. The sliver does not return to the geni_provisioned state on sliver expiration.

All slivers which began in the geni_allocated state will remain in the geni_allocated state, or will have been deleted and there will be no way (except for attempting another Update()) to revert their configuration. Slivers that are updated and remain in the geni_allocated state will have a new expiration time, as determined by the aggregate, but typically set just as for new slivers returned by Allocate(). As with other geni_allocated slivers, if the sliver expires before it is Provision()ed, then the resources are freed and the sliver is geni_unallocated.

Any slivers which began in the geni_updating state will remain in the geni_updating state, with their prior existing expiration time. If the expiration time is reached before the sliver is Provision()ed, then the resources are freed and the sliver is geni_unallocated. The sliver does not return to the geni_provisioned state on sliver expiration.

Any new resources requested in the RSpec not associated with a current sliver will be in the geni_allocated state and will be added to slivers, just as after a call to Allocate().

While slivers are geni_updating, operational actions are not permitted (PerformOperationalAction() should return an error code - 2, 14, or 16).

Return

On success, the value field of the return struct will contain a struct:

{
 geni_rspec: <geni.rspec manifest of newly allocated slivers>,
 geni_slivers: [
        {
                  geni_sliver_urn: <string sliver urn>
                  geni_expires: <dateTime.rfc3339 allocation expiration string, as in geni_expires from Status>,
                  geni_allocation_status: <string sliver state - e.g. geni_allocated, geni_updating>,
                  geni_next_allocation_status: <string state of the sliver after the next Provision() call - geni_unallocated or geni_provisioned>,
                  geni_error: <optional string, may be omitted entirely, explaining any failure for a sliver>
        },
        ...
    ]
}

The list of returned slivers includes even those that will be deleted. For those slivers, geni_next_allocation_status will be geni_unallocated.

State Guarantees

The manifest RSpec may include an extension describing what guarantees the AM will make about state preservation. Omitting the extension means the AM provides no guarantees. It is proposed that the extension have the following RNC schema:

default namespace = "http://www.geni.net/resources/rspec/ext/preserve/1"

# This is meant to extend a node or link
Preserved = element preserve {
  attribute guarantee { "geni_none" | "geni_persistent_state" |
                        "geni_dynamic_state" | "geni_no_disruption" }
}

start = Preserved

Guarantee levels are defined to mean:

  • geni_none is returned when the AM provides no state preservation guarantees for a sliver.
  • geni_persistent_state is returned when the AM may wipe out dynamic state (i.e. reboot a node) but will preserve persistent state.
  • geni_dynamic_state is returned when the AM retains the dynamic state but there may be other perturbations such as packet loss or service restarts.
  • geni_no_disruption is returned when the AM will not make any changes that can perturb the sliver.

Cancel()

Cancels an Update() or an Allocate().

struct Cancel(string urns[], string credentials[], struct options)

This call accepts the geni_best_effort option; as with Delete(), when this option is supplied the aggregate should attempt to de-allocate individual slivers.

When applied to slivers in the geni_allocated state, Cancel() acts just like Delete(). When applied to slivers in the geni_updating state, those slivers are returned to the geni_provisioned state with no change to the operational state or attributes of the existing slivers.

Return: On success, the value field of the return struct will contain a struct:

{
   geni_rspec: <geni.rspec, a Manifest RSpec>
   geni_urn: <string slice urn of the containing slice>
   geni_slivers: [
               {
                  geni_sliver_urn: <string sliver urn>
                  geni_expires: <dateTime.rfc3339 allocation expiration string, as in geni_expires from SliversStatus>,
                  geni_allocation_status: <string sliver state - e.g. geni_allocated or geni_provisioned >,
                  geni_operational_status: <string sliver operational state>,
                  geni_error: <optional string, may be omitted entirely, explaining any failure for a sliver>
               },
               ...
         ]
}

Changes to Provision()

When applied to slivers in the geni_updating state, this call will either remove those slivers (if they were marked for deletion as described in the Update() call) or move them to the geni_provisioned state, provisioning them according to the manifest RSpec returned from that call.

Changes to Describe()

The addition of an Update() operation creates an ambiguity on the slivers being described. Should the Aggregate Describe() the slivers as they will be instantiated on a successful Provision() (including all slivers that are geni_allocated, geni_updating, and geni_provisioned)? Or should the AM describe the slivers which would remain after a Cancel() (including only the slivers that are geni_allocated or geni_provisioned, plus the updated slivers as they existed before being updated)?

In order to account for this ambiguity, Describe() accepts a new option, geni_cancelled which defaults to false. When geni_cancelled is true, Describe() returns the state of all currently geni_provisioned slivers, showing the user what their slice will look like if they cancel all geni_updating and geni_allocated slivers. When geni_cancelled is false, the manifest and status will reflect the changed slivers in the geni_updating and geni_allocated states.

Additionally, the per-sliver return structure for Describe() adds a new attribute, just as for Update():

                  geni_next_allocation_status: <string state of the sliver after the next Provision() call - geni_unallocated or geni_provisioned. May be an empty string.>

This attribute is an empty string (not None or null) when the current allocation state is not geni_allocated or geni_updating.

Changes to Delete()

A call to Delete() can be used to delete slivers in any state.

Changes to PerformOperationalAction()

While slivers are geni_updating, operational actions are not permitted (PerformOperationalAction() should return an error code - 2, 14, or 16).

Changes to Status()

The per-sliver return structure for Status() adds a new attribute, just as for Update():

                  geni_next_allocation_status: <string state of the sliver after the next Provision() call - geni_unallocated or geni_provisioned. May be an empty string.>

This attribute is an empty string (not None or null) when the current allocation state is not geni_allocated or geni_updating.

Changes to Renew()

The per-sliver return structure for Renew() adds a new attribute, just as for Update():

                  geni_next_allocation_status: <string state of the sliver after the next Provision() call - geni_unallocated or geni_provisioned. May be an empty string.>

This attribute is an empty string (not None or null) when the current allocation state is not geni_allocated or geni_updating.

Change to geni_single_allocation

When an AM reports geni_single_allocation as true in GetVersion(), the following additional restrictions apply to calls:

  • When calling Update(), the urns list must either contain a slice URN or it must contain the URN of every sliver at this Aggregate Manager in the given slice.
  • When calling Provision(), the urns list must contain the sliver URNs of all slivers in either the geni_allocated or geni_updating states in the given slice.
  • When calling Cancel(), the urns list must contain the sliver URNs of all slivers in either the geni_allocated or geni_updating states in the given slice.

Mixed State Calls

A Renew() call which contains sliver URNs in multiple allocation states will attempt to renew them all to the same requested expiration time.

A Delete() call which contains sliver URNs in multiple allocation states will attempt to delete all slivers and put them in the geni_unallocated state.

Provision() and Cancel() will ignore any sliver URNs which are not in the geni_allocated or geni_updating states.

Update() can be called on slivers in any state. It will have no effect on slivers in the geni_unallocated state, and otherwise operates as described above.

RSpec Change

To easily accommodate submitting Update() requests, RSpecs will be changed:

  • sliver_id will be allowed in request RSpecs
  • Manifests will be modified to be re-usable as request RSpecs. In particular, the manifest schema will become a simple extension of the request schema, but this change is not required to support Update().

Change Set N: Add information to GetVersion

This change set adds some additional information to the return from GetVersion.

This change set was discussed at the GEC15 coding sprint and adopted at GEC19. Aggregates may implement this as part of their AM API v3 implementation, but will be required to do so as part of AM API v4.

Change Set N1: Add geni_am_code_version

Aggregates use varying ways to identify the software version their AM uses. PlanetLab and ProtoGENI use code_tag, Orca uses orca_version, and FOAM uses foam_version. Automatically identifying software versions is important for monitoring and GENI operations. Aggregates are therefore strongly encouraged to advertise their current software revision using this standard field, though for security reasons some aggregates may choose not to do so.

This change set proposes adding geni_am_code_version to the return from GetVersion.

  • The field is optional. Aggregate developers are expected to include this option, but site operators may select not to expose it.
  • The value is a string.
  • Legal characters are: alphanumeric, space, - (hyphen), ., : (colon), #, _ (underscore), +, (, )
  • Regular expression: '^[a-zA-Z0-9-\.:#_\+\(\)]+$'

Change Set N2: Add geni_am_type

Aggregates are allowed to include AM specific return values, or even implement custom methods as part of the same interface.

Currently, there is no consistent way to identify what type of aggregate a given instance is, so as to know what additional attributes or options may be used. This proposal would add a way for aggregates to identify what kind of aggregate this is, and therefore what aggregate specific options or returns are applicable. Aggregates of aggregates may identify as multiple types. One of those types would indicate that this is such an aggregate of aggregates, and other listed types would indicate that clients may interact with the aggregate as though it is any of the listed types.

This change set proposes adding geni_am_type to the return from GetVersion.

  • The field is required.
  • The value is a list of strings, of length at least one. It should generally be a list of length 1. Aggregates of aggregates may list multiple types.
  • Legal characters: alphanumeric
  • Values should be one of the defined GENI AM types if applicable, as defined by the AM API geni-am-types.xml. (As of this proposal, one of orca, foam, protogeni, sfa, dcn. More GENI AM types may be added in the future.)

Change Set N3: Use consistent types and prefixes

Make am_type, geni_am_type, and AM specific prefixes for new options consistent.

The AM API specifies an am_type field as part of the code in the basic return structure. This field has a set of prescribed values.

An earlier proposal adds a geni_am_type return to GetVersion.

The AM API encourages AM specific options and returns. These options and returns should have unique names, but the API does not specify how to ensure this.

This proposal aims to standardize these fields.

am_type and geni_am_type:

  • The value is a string.
  • Legal characters: alphanumeric
  • Values should be one of the defined GENI AM types if applicable, as defined by the AM API geni-am-types.xml. (As of this proposal, one of orca, foam, protogeni, sfa, dcn. More GENI AM types may be added in the future.)

Aggregate specific options and return attributes should be named with a prefix denoting the aggregate type. It should be the same as the am_type and geni_am_type for this aggregate. Therefore, character restrictions are:

  • The value is a string.
  • Legal characters: alphanumeric
  • Values should be one of the defined GENI AM types if applicable, as defined by the AM API geni-am-types.xml. (As of this proposal, one of orca, foam, protogeni, sfa, dcn. More GENI AM types may be added in the future.)

For example: ProtoGENI includes an AM-specific status in the v2 API return from SliverStatus called pg_status. By this proposal, that attribute should be renamed protogeni_status.

Change Set O: Refine character restrictions

This change set was dicussed at the GEC15 coding sprint. Aggregates may implement this as part of their AM API v3 implementation, but will be required to do so as part of AM API v4.

Change Set O1: Allow other characters in sliver names

Currently we heavily constrain legal characters in the 'name' portion of sliver URNs. See GeniApiIdentifiers. This is awkward. In particular characters like periods, and underscores are not allowed, but are very useful as separators.

This change set was adopted at GEC19. Aggregates may start using these modified sliver names.

This change set proposes modifying the name portion of the URN rules as follows. Note that these changes loosen existing restrictions, and so are backwards compatible (existing sliver names remain legal). Note also that this change simple allows aggregates to use these other characters in their own identifiers, if they so choose.

Sliver:

  • May use only alphanumeric characters plus hyphen, underscore, or period: '^[a-zA-Z0-9-_\.]+$'

Change Set O2: Explicitly define legal characters in some strings

Various fields in the AM API specification are defined only as strings, without explicit limits on legal characters. This makes it awkward to pass these values to and from scripts. In particular, many fields should only be alphanumeric, plus some limited number of separator characters.

This change set was adopted at GEC19. This change is required in AM APIv4, optional earlier.

This change set proposes constraining several options/return values that are currently defined only as 'string'.

  1. From getversion and listresources and describe, the RSpec type:

May use only alphanumeric characters plus hyphen, underscore, period, or colon; alphanumeric only in first character. '^[a-zA-Z0-9][a-zA-Z0-9-_\.:]*$'

  1. From getversion and listresources and describe, the RSpec version:

May use only alphanumeric characters plus hyphen, underscore, period, or colon; alphanumeric only in first character. '^[a-zA-Z0-9][a-zA-Z0-9-_\.:]*$'

  1. From getversion, the RSpec schema, and namespace:

These are standard XML schemas and namespaces, and should follow the applicable standards. See http://www.w3.org/TR/xml-names11/ and http://www.w3.org/TR/xmlschema11-1/

  1. From getversion and the credentials argument, the credential geni_type:

May use only alphanumeric characters plus hyphen, underscore, period, or colon; alphanumeric only in first character. '^[a-zA-Z0-9][a-zA-Z0-9-_\.:]*$'

  1. From the return of multiple methods, geni_operational_status: alphanumeric plus underscore; alphanumeric only in first character: '^[a-zA-Z0-9][a-zA-Z0-9_]*$'

Change Set P: Support proxy clients that 'Speak For' an experimenter

This change set was adopted at GEC19. This functionality may be implemented in AM API v2, and is required in AM API v4.

GENI tools invoke AM API methods on behalf of experimenters. When that tool instance is well known and trusted by the experimenter, and runs local to a single experimenter, then it is reasonable for that tool to be given the experimenter's private key and public certificate, and 'speak as' the experimenter. But for a hosted tool that might act for multiple experimenters it is both more secure and more clear what is happening for such tools to have their own identity certificate and private key by which each tool instance authenticates to AM API calls, and use a new 'Speaks For' credential (or credential set) to authorize the tool instance to take a given action. The experimenter will issue such credentials to the tool instance (possibly scoped by time, slice, aggregate, or other dimension), and the aggregate can then properly assign the resources to the experimenter, while logging and reporting that the tool performed the operation. This functionality can also be used for an experimenter to authorize a 'proxy' aggregate or an aggregate of aggregates.

Two changes are required in the AM API to support this functionality. First, not all aggregates will support 'Speaks For' yet, so tools must be able to distinguish. Second, we require an explicit way for the aggregate to determine on whose behalf a tool is taking this action.

  1. Add a new optional return entry from GetVersion:

geni_handles_speaksfor: A boolean (0 or 1 in XML-RPC), default is false (0).

When present and true, the aggregate is capable of authorizing AM API actions using a GENI 'Speaks For' credential (or credential set), and accepting the option geni_speaking_for (see below). When the aggregate supports 'Speaks For', tools should authenticate using their own GENI issued certificate and key, supply a 'Speaks For' credential (or credential set) from the experimenter authorizing the tool instance to perform the given action, and supply the URN of the experimenter on whose behalf the tool is acting in the geni_speaking_for option. When the aggregate does not support 'Speaks For', the tool is expected to 'speak as' the experimenter, using the experimenter's own GENI issued certificate and key. Aggregates should however attempt to support 'Speaks For', as 'Speaks As' is undesirable deprecated behavior. If a client supplies 'Speaks For' arguments to an aggregate that does not advertise support for 'Speaks For', the aggregate will typically ignore these, and attempt to authorize the client directly (though that behavior is not required), likely resulting in an authorization failure.

  1. Add a new option to the options field of each AM API call:

geni_speaking_for. <string URN>

When supplied, this is the URN of the experimenter on whose behalf the client is acting. The aggregate should expect to find a 'Speaks For' credential or credential set in the credentials argument, by which the given experimenter authorizes the client (as authenticated by the SSL client certificate) to perform the given action on their behalf. If such a credential is not found, the aggregate should fail the request with an authorization error (e.g. error code 3: FORBIDDEN). When omitted, the aggregate need not look for a 'Speaks For' credential, but should look to authorize the client itself ('speaks as'). If the client is not itself directly authorized to perform the given action, the aggregate should fail the request with an authorization error (e.g. error code 3: FORBIDDEN).

Two other additions are required: we must define the 'Speaks For' credential and its semantics, and we must define the URN and certificates for tools.

The 'Speaks For' credential will be a signed XML document encoding of an ABAC credential (GENI type geni_abac version 1) as specified here. Several general points are worth noting:

  • The credential includes the certificate of the user and an identifier for the tool. For the credential to be accepted, each certificate must itself be trusted by the aggregate; current recommended GENI policy requires a particular format for that certificate, and requires that it be signed (directly or indirectly) by a trusted GENI root.
  • The credential includes an expiration
  • The credential may in future include scope limitations (including slice, aggregate, operation)
  • When the aggregate authorizes a 'Speaks For' operation, the aggregate must treat the operation as though performed by the experimenter, but also log that it was done via the given tool. That is, resources will be owned by the experimenter, and logs and monitoring reports will include both the experimenter URN and the tool URN.

The specific ABAC assertion that this 'Speaks For' credential encodes is: Experimenter.speaks_for_Experimenter <- Tool. That is, the experimenter signs an assertion saying that the Tool has the attribute in the Experimenter's namespace speaks_for_Experimenter, where the Experimenter and Tool are identified (as with all ABAC assertions) by the SHA1 hash of their public key. As an example, here is a selection from a 'Speaks For' credential:

    <expires>2014-02-12T20:10:32Z</expires>
    <abac>
      <rt0>
        <version>1.1</version>
        <head>
          <ABACprincipal>
            <keyid>097f010966eacbc0f8e2fc8c66c8abfdd55f6036</keyid>
          </ABACprincipal>
          <role>speaks_for_097f010966eacbc0f8e2fc8c66c8abfdd55f6036</role>
        </head>
        <tail>
          <ABACprincipal>
            <keyid>66dd9f3018e64c12130068f4a71d364fc9cbdfb6</keyid>
          </ABACprincipal>
        </tail>
      </rt0>
    </abac>

Tool Certificates

Tool instances may be issued GENI identity certificates, using the same format and rules as for users.

  • The URN will be the URN of the tool instance. With this change, we introduce a new 'type' for the URN field: tool. The name of a tool is subject to the same restrictions as the name for users:
    • Tool names are case-insensitive internally, though they may be case-sensitive in display.
      • EG JohnSmth as a display name is johnsmth internally, and there cannot also be a user JOHNSMTH.
    • Tool names should begin with a letter and be alphanumeric or underscore, hyphen, at sign or period: ('^[a-zA-Z][a-ZA-Z0-9\-_@\.]{0,63}$').
    • Tool names are limited to 64 characters.
    • Tool URNs (which contain the authority name and the tool instance name) are required to be temporally and globally unique.
    • Tool names should encode both the tool type and instance. For example portal-gpo or genidesktop-uky.
  • The tool email address should be a way to contact the administrators of the tool instance - the organization or individual who applied for the certificate and who stands behind its integrity.

Note that this functionality also supports proxy aggregates, or aggregates of aggregates. See a previous related change proposal: http://groups.geni.net/geni/wiki/GAPI_AM_API_DRAFT?version=47#ChangeSetJ:Proxyaggregatemanagersaresupported

Change Set Q: Support changing users and keys on existing compute slivers

This change set was adopted at GEC19.

Experimenters often want to change the users and/or SSH keys installed on existing running compute nodes, in a way that is persistent and consistent with any aggregate manager controlled processes. This is theoretically possible through the use of Update(), but the semantics are not clear there.

This change set proposes a way to do this using PerformOperationalAction() and also proposes the semantics of providing a new set of users/keys to the Provision() call after a call to Update().

This change set may be implemented as part of an aggregate's AM API v2 implementation, not just the AM API v3 implementation

Specifying users and keys

Users to be installed and SSH keys to be installed for those users are specified using the geni_users[] struct as specified in AM API v3.

When applied to an existing sliver (through a new operational action or through a call to Update() and Provision()):

  • This same user struct is to be used
  • New users may be added
  • Omitting a user means there will be no change to the keys installed for that user. It does not mean the user should be removed.
  • Existing users will have all SSH keys replaced by the new set. Note that a user may have an empty list of SSH keys specified, effectively preventing the user from accessing the node.

Semantics with Update

After a call to Update(), users must call Provision() to commit the change. Provision() takes the geni_users[] struct. When called after Update(), the geni_users[] struct will have the add/replace semantics described above for any existing nodes; new nodes will treat all the users/keys as additions as normal.

Changing users/keys through an operational action

Users and SSH keys can be changed using 2 new operational actions (through PerformOperationalAction()):

  1. geni_update_users: The credentials[] argument must include credentials over the slice as usual. The options struct must include the geni_users option as specified in AM API v3 and with the semantics described above. This action is only legal on slivers in the geni_ready operational state. This action immediately moves all such slivers to a new geni_updating_users operational state. Slivers stays in that state until the aggregate completes the needed changes, at which time the slivers change back to the geni_ready operational state. Slivers may be in the geni_updating_users state for several minutes; during this time no other operational actions can be taken on the slivers.

Besides the usual return error codes, the PerformOperationalAction method with this action may return REFUSED if the sliver is in the wrong operational state.

Note that as described here, aggregates supporting this action and operational state will advertise this fact in their advertisement RSpec.

Sample JSON value for the geni_users option to be supplied:

{
 "geni_users": [
  {
   "urn": "urn:publicid:IDN+ch.geni.net+user+jdoe",
   "keys": ["ssh-rsa jdoegoodkey"]
  },
  {
   "urn": "urn:publicid:IDN+ch.geni.net+user+jsmith",
   "keys": ["somekey", "someotherkey"]
  }
 ]
}
  1. geni_updating_users_cancel: This action requires no options. It cancels any pending geni_update_users action on the named slivers, returning those slivers to the geni_ready operational state. This action is only legal on slivers in the geni_updating_users operational state. This action may be used on slivers which fail to complete the geni_update_users action. After a successful geni_updating_users_cancel, the state of users and keys on the sliver(s) is not defined; some may have the same users/keys as they had prior to beginning the geni_update_users action, and others may have already updated to the new set of users and keys.

Besides the usual return error codes, the PerformOperationalAction method with this action may return REFUSED if the sliver is in the wrong operational state.

Change Set R: Allow Renew to extend slivers as much as possible

Adopted at the GEC 18 Coding Sprint. This proposal was discussed at the GEC 18 Coding Sprint. ProtoGENI/InstaGENI, ION/MAX, Orca/ExoGENI, and FOAM all agreed to implement and support this option.

In the current AM API, aggregates are free to restrict how long a reservation can be extended with local policy. AM API v2 says the aggregate should return false if the reservation cannot be extended as requested. V3 allows the geni_best_effort option, which would allow some slivers to be extended, while other requests failed. But it still does not permit an aggregate to extend a reservation for less than the full request. This makes things difficult for experimenters.

Currently, if an experiment requests a renewal past the aggregate's local policy maximum, the request fails - and there is no standard way to determine from the error message in v3 (you may get none in v2) what renewal time 'would' work. The problem is compounded if you try to renew slivers at multiple aggregates.

This proposed change set would accommodate an experimenter attempting to extend their reservation as long as possible

Proposal: Add a new option: geni_extend_alap (as long as possible), default is False (XML-RPC boolean) (preserving current behavior). Option is valid in AM API v2 in RenewSliver, or in v3 or higher in Renew. AMs are not required to support it. When True, the AM should extend the life of the slice/sliver (depending on the API version), to the minimum of the requested new expiration time, the expiration time in the slice credential (of course), and the AM local maximum based on local policy (which may be resource or slice dependent).

When the AM successfully extends the life of the slice/sliver using this new option:

  • return value is success (True or the API v3 struct)
  • AM API v3 Renew return struct will be as previously defined, including the new sliver expiration time
  • For AM API v2 aggregates, the output slot of the AM API triple should be the new sliver expiration time - either the time requested, or the 'as long as possible' time. This expiration time should be a standard GENI AM API datetime complying with RFC 3339.

When the AM cannot extend the life of the slice/sliver at all (because it does not honor the new option, or because no expiration is possible), then it fails the request (return False in APIv2, return an error code in APIv3 unless geni_best_effort is True, in which case it may return success and indicate the individual per-sliver errors in the return struct; v3 error code may be one of UNSUPPORTED for an AM that does not support this option, or BADARGS or ERROR for other failures).

Note that this would be a new option that AMs are not required to honor - making things backwards compatible.

Comments

  • Maybe we should try to convey AM local policy on max renewal times (which may be per resource type) in GetVersion or the Ad RSpec ('before' you try to reserve or renew the resources). That is a separate discussion.

Attachments (1)

Download all attachments as: .zip