Cache the AAI data and refresh them periodically so that Holmes won't have to make an HTTP call to AAI every time it tries to correlate one alarm to another.
The problem for caching is how to know when to update the cached data. Even though the access time may be fast for Holmes, the risk is using out-of-date data, so the correlations will be wrong anyway. Also, duplicating the AAI data outside of AAI is probably a bad architectural decision. Making AAI faster for these use cases would be better.
Has there been a performance analysis of where the time is spent? Could it help to use ElasticSearch (e.g. as in sparky)? Should Holmes have a batch interface to get more AAI data in fewer calls? Or a better correlation API that results in fewer calls?
Guangrong Fu will try custom queries for queries that took to long to return
The hardware (mainly storage) influences the query speed - need to find out what hardware was the speed test conducted on (Guangrong Fu will provide HW specs)
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
HOLMES-186
Would the AAI Cacher
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
AAI-1337
help to improve performance?
5th Mar: Guangrong Fu
Hi,
Sorry for my late response. It took me a long time to set up AAI in my own env. For Item 10, here's some information:
Main APIs invoked in Holmes for different use cases:
VoLTE
Getting the VM query URL via: /search/nodes-query?search-node-type=vserver&filter=vserver-name:EQUALS: - once
Getting VM info via: the URL returned by the query above - once
Getting the VNF data via: network/generic-vnfs/generic-vnf - once
CCVPN
Updating terminal point via: /network/pnfs/pnf/{pnfName}/p-interfaces/p-interface/nodeId-{pnfName}-ltpId-{ifName} - once
Getting logical links via: /network/pnfs/pnf/{pnfName}/p-interfaces/p-interface/nodeId-{pnfName}-ltpId-{ifName} - 3 times
Getting VPN bingding info via: /network/pnfs/pnf/{pnfName}/p-interfaces/p-interface/nodeId-{pnfName}-ltpId-{ifName} - once
Getting connectivity info via: /network/vpn-bindings/vpn-binding/{vpnId} - once
Getting service instance info via: /network/connectivities/connectivity/{connectivityId} - once
Performance
We set up an AAI env on a VM (8 cores, 16GB memory, 160GB storage) following the guidance https://wiki.onap.org/display/DW/How+to+Docker+setup+on+Single+VM+HEAT+Deployment and tried to run a VNF query using "/aai/v11/cloud-infrastructure/cloud-regions/cloud-region/example-cloud-owner-val-45051/example-cloud-region-id-val-56689/tenants/tenant/example-tenant-id-val-51834/vservers/vserver/example-vserver-id-val-51834" (which is returned by "/search/nodes-query?search-node-type=vserver&filter=vserver-name:EQUALS:") for 1000 times. It took ~95ms per query. Also, we tried to query a VNF for 1000 times via "/aai/v11/network/generic-vnfs/generic-vnf/example-vnf-id-val-92494" and the average time is ~86ms.
From the result, we know that even for a single request, the time cost reaches around 100ms. Let alone there will be several requests sent to AAI when an alarm is processed by Holmes. Taking CCVPN for example, for each alarm, there are up to 7 requests made. That means it'll take around 600-700 ms for Holmes to interact with AAI. In case of alarm storms, it is hard for AAI to support such intensive queries.
6th March: Guangrong Fu
In my opinion, the performance of AAI queries is not only impacted by the computation inside AAI, but also impacted by the HTTP request itself.
I've done another test. I tried to send requests to the health check API (which does nothing but return immediately after it receives a request ) of Holmes. The average time cost is also ~ 70ms. So it seems to be a problem with the time cost caused by setting up and releasing HTTP connections.
6th March: Keong
Regarding these queries:
Getting logical links via: /network/pnfs/pnf/{pnfName}/p-interfaces/p-interface/nodeId-{pnfName}-ltpId-{ifName} - 3 times
Getting VPN bingding info via: /network/pnfs/pnf/{pnfName}/p-interfaces/p-interface/nodeId-{pnfName}-ltpId-{ifName} - once
What depth is used on these GET calls? If the defaulting to depth=0, then perhaps some improvement can be made by using "depth=1" or "depth=2"? Fewer calls returning more data could improve overall performance.
Same could be achieved by changing to Nodes query, e.g.
GET /aai/v14/nodes/p-interfaces?interface-name=nodeId-{pnfName}-ltpId-{ifName}
Question2: Would it help to have the Holmes pod co-located with the AAI haproxy and AAI resources pods? Reduced network latency could improve overall performance.
Guangrong: Holmes is acutally deployed by DCAE. I'm not sure whether your proposal is feasible. What's more, the performance data I got was based on the fact that Holmes and AAI were deployed on the same VM, sharing the same docker env.
how to minimise impact of the transition from pnf-name as unique to pnf-id as unique key?
would the v14 URL be different from the v15 URL? would both paths be equally supported for GET/PUT/etc?
what forwards-compatibility or backwards-compatibility will be supported?
how to migrate forwards or backwards database versions, ONAP versions, etc, across this transition?
who is going to implement it? Test it?
what is the impact of this not going ahead?
William LaMont will check for existing migration utility that handles this use case (changing the key from one existing attribute to another)
James Forsyth will socialize the breaking change on the PNF in the next PTL call so clients can prepare to do a search for ?pnf-name=${pnf-name} instead of /pnfs/pnf/${pnf-name}. They also need to handle doing the PUT operation differently - Added to PTL agenda PTL 2019-02-19
In think it would be good to answer what is the meaning of the field (collection of PEMs of the CA xor URL)
Questions:
1. Is AAI intended to strictly prescribe how the fields are used and what contents are in the values? 2. Or does AAI simply reflect the wishes of all the client projects that use it to store and retrieve data?
Even if (1) is true, AAI is not really in any position to enforce how clients use the data, so really (2) is always true and we need to consult the original producers of the data and the ultimate consumers of the data to document their intended meanings.
How do we push to have documentation on the purpose and meaning of the fields in AAI?
Where does all this documentation go?
Should the documentation be backed up by validation code?
The below put_vLB.sh script can be used to submit the vLB data to A&AI in order to run ConfigScaleOut use case. This script and referenced JSON files are used on an AAI instance where the cloud-region and tenant are already defined.
Similarly:
Jira
server
ONAP JIRA
Regarding
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
OOM-1652
/
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
OOM-1676
/
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
OOM-1186
/
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
OOM-1187
/
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
OOM-1194
/
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
TEST
OOM-
133
1199
:
Is this something AAI team needs to be aware of? Or is it OK for OOM team to just switch it around?
Will this introduce unexpected dependencies between AAI, AAF, OOF, Portal and SDC that will create difficulties for upgrade/downgrade/backup/restore/maintenance/schema change?
Is it going to exacerbate the performance problems already noted?
The "rolling upgrade" change has been combined with the "shared cassandra" change.
The "shared cassandra" change has been combined with the "AAF shared cassandra" change, which means it's also combined with the "AAI shared cassandra" change.
Use MUSIC or any other alternative in memory caching like Redis etc?
Optimize flavor retrieval from A&AI and Cache the information if necessary
See also
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
OPTFRA-268
/
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
OPTFRAAAI-291
Similarly to the "AAI too slow for Holmes" item below, this introduction of extra caching of AAI data is a worrisome development and sad indictment of the performance of the system architecture.
What can we do about this?
Would the AAI Cacher
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
AAI-1337
help to improve performance?
10
MultiCloud usage of AAI for HPA telemetry/time-series data to OOF
In think it would be good to answer what is the meaning of the field (collection of PEMs of the CA xor URL)
Questions:
1. Is AAI intended to strictly prescribe how the fields are used and what contents are in the values? 2. Or does AAI simply reflect the wishes of all the client projects that use it to store and retrieve data?
Even if (1) is true, AAI is not really in any position to enforce how clients use the data, so really (2) is always true and we need to consult the original producers of the data and the ultimate consumers of the data to document their intended meanings.
How do we push to have documentation on the purpose and meaning of the fields in AAI?
Where does all this documentation go?
Should the documentation be backed up by validation code?
HPA telemetry data collection and make it persistent in A&AI, from which OOF can leverage during its decision making process.
and
1. Multi-cloud to collect the data from time-series data services like Prometheus (http://prometheus.io) or openstack Gnocchi, and push them to A&AI based on the data recording & aggregation rules.
and
The reason why we propose here is that VES mechanism doesn't store the telemetry data into A&AI. And OOF now can only get those kind of data from A&AI.
Some concerns:
how much additional load will this place on AAI?
will AAI cope with this load?
is AAI suitable for "time-series data"?
is "telemetry data" considered to be "active & available inventory"?
should OOF access the telemetry/time-series data via other means (not AAI)?
AAI API latency (4~6 second per request as benchmarked in CMCC lab) could be a problem
I agree & acknowledge that atleast for a foreseeable future, we would need a way to specify the VNFM / NFVO as part of "Design Decision", which I believe can be reflected as part of VNFD/NSD ( using some special attribute ) or as part of internal Model that SDC might build before distributing the same. SO can then use this hint to select relevant actors. My only question is, why this has to be maintained in AAI which is exclusively for runtime record? All AAI cares about is what is running in the network irrespective of how that got orchestrated. Isn't it ?
On a broader note, I would like to understand what's the original intent of AAI ( atleast in ECOMP world ) ? Are we simply assuming that, just because AAI has "available inventory" in its name, we are expecting it to keep track of cloud inventory realtime ? Because our entire story ( including the new G-FPS proposal ) is based on this assumption. Can anyone from AAI team or ATT clarify on this ?
Because AFAIK, AAI neither has the schema to host such available inventory, nor the MC has the pub/sub or polling mechanism ( today ) to refresh the cloud inventory inplace. Ofcourse those can be scoped for further releases, but my original question is, was that the original intent behind AAI or are we now including it in the scope?
For the first question: I think that A&AI needs to maintain the VNF instance ↔ VNFM instance and the NS instance ↔ NFVO instance relationship for subsequent life cycle operations, i.e. a scale or heal operation. The path would be something like Event (VNF Instance, Busy) → DCAE (policy for VNF instance) → Policy Evaluation (VNF instance, Scale-out) → SO (VNF instance, Scale-out) → A&AI (find VNFM instance for the VNF instance) → SO (VNF instance, VNFM instance, Scale-out) → SOL003 Adapter (VNFM instance, VNF instance, Scale-out) → VNFM instance (VNF Instance, Scale-out).
As I understand, ESR has "esr-vnfm-list", which has an "esr-vnfm", which has "esr-system-info-list", which has "esr-system-info", which has a "relationship-list" that can contain relationships to "generic-vnf" and other AAI objects.
The "generic-vnf" object also contains "self-link", "ipv4OamAddress", "ipv4OamGatewayAddress", etc, which links the AAI object back to its "source-of-truth" external-system.
Is there some new data, new schema or new API that is required on top of this?
Fernando Oliveira; Apologies for my lack of knowledge, but a few comments:
For the VNF/VF ↔ VNFM case, I think that there needs to be a reference from a VNF/VF instance record to the specific instance of the VNFM that was used to deploy the VNF/VF. If there is already such a reference from the VNF/VF through the ESR to the specific item on the esr-vnfm-list, then I think that would be sufficient. If not, I think that would be a new requirement.
For the Service ↔ NFVO case, Is there an equivalent NFVO/Orchestrator list in the ESR? The esr-nfvo-list would need the same set of info as the VNFM case. If the esr-nfvo-list does exist, I think that there needs to be a reference from the Service Instance record to the specific NFVO instance that deployed the Service. Is there such a reference? If not, I think that would be a new requirement.
Bo Lv can comment more on the current ESR capabilities, but I believe there are only 3 kinds of systems so far: EMS, VNFM and third-party SDNC.
ESR could be extended to handle VNFO as another kind of system.
Is this something AAI team needs to be aware of? Or is it OK for OOM team to just switch it around?
Will this introduce unexpected dependencies between AAI, AAF, OOF, Portal and SDC that will create difficulties for upgrade/downgrade/backup/restore/maintenance/schema change?
Is it going to exacerbate the performance problems already noted?
The below put_vLB.sh script can be used to submit the vLB data to A&AI in order to run ConfigScaleOut use case. This script and referenced JSON files are used on an AAI instance where the cloud-region and tenant are already defined.
Use MUSIC or any other alternative in memory caching like Redis etc?
Optimize flavor retrieval from A&AI and Cache the information if necessary
See also
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
OPTFRA-268
/
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
OPTFRA-291
Similarly to the "AAI too slow for Holmes" item below, this introduction of extra caching of AAI data is a worrisome development and sad indictment of the performance of the system architecture.
What can we do about this?
Would the AAI Cacher
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
AAI-1337
help to improve performance?
12
MultiCloud usage of AAI for HPA telemetry/time-series data to OOF
HPA telemetry data collection and make it persistent in A&AI, from which OOF can leverage during its decision making process.
and
1. Multi-cloud to collect the data from time-series data services like Prometheus (http://prometheus.io) or openstack Gnocchi, and push them to A&AI based on the data recording & aggregation rules.
and
The reason why we propose here is that VES mechanism doesn't store the telemetry data into A&AI. And OOF now can only get those kind of data from A&AI.
Some concerns:
how much additional load will this place on AAI?
will AAI cope with this load?
is AAI suitable for "time-series data"?
is "telemetry data" considered to be "active & available inventory"?
should OOF access the telemetry/time-series data via other means (not AAI)?
AAI API latency (4~6 second per request as benchmarked in CMCC lab) could be a problem
I agree & acknowledge that atleast for a foreseeable future, we would need a way to specify the VNFM / NFVO as part of "Design Decision", which I believe can be reflected as part of VNFD/NSD ( using some special attribute ) or as part of internal Model that SDC might build before distributing the same. SO can then use this hint to select relevant actors. My only question is, why this has to be maintained in AAI which is exclusively for runtime record? All AAI cares about is what is running in the network irrespective of how that got orchestrated. Isn't it ?
On a broader note, I would like to understand what's the original intent of AAI ( atleast in ECOMP world ) ? Are we simply assuming that, just because AAI has "available inventory" in its name, we are expecting it to keep track of cloud inventory realtime ? Because our entire story ( including the new G-FPS proposal ) is based on this assumption. Can anyone from AAI team or ATT clarify on this ?
Because AFAIK, AAI neither has the schema to host such available inventory, nor the MC has the pub/sub or polling mechanism ( today ) to refresh the cloud inventory inplace. Ofcourse those can be scoped for further releases, but my original question is, was that the original intent behind AAI or are we now including it in the scope?
For the first question: I think that A&AI needs to maintain the VNF instance ↔ VNFM instance and the NS instance ↔ NFVO instance relationship for subsequent life cycle operations, i.e. a scale or heal operation. The path would be something like Event (VNF Instance, Busy) → DCAE (policy for VNF instance) → Policy Evaluation (VNF instance, Scale-out) → SO (VNF instance, Scale-out) → A&AI (find VNFM instance for the VNF instance) → SO (VNF instance, VNFM instance, Scale-out) → SOL003 Adapter (VNFM instance, VNF instance, Scale-out) → VNFM instance (VNF Instance, Scale-out).
As I understand, ESR has "esr-vnfm-list", which has an "esr-vnfm", which has "esr-system-info-list", which has "esr-system-info", which has a "relationship-list" that can contain relationships to "generic-vnf" and other AAI objects.
The "generic-vnf" object also contains "self-link", "ipv4OamAddress", "ipv4OamGatewayAddress", etc, which links the AAI object back to its "source-of-truth" external-system.
Is there some new data, new schema or new API that is required on top of this?
Fernando Oliveira; Apologies for my lack of knowledge, but a few comments:
For the VNF/VF ↔ VNFM case, I think that there needs to be a reference from a VNF/VF instance record to the specific instance of the VNFM that was used to deploy the VNF/VF. If there is already such a reference from the VNF/VF through the ESR to the specific item on the esr-vnfm-list, then I think that would be sufficient. If not, I think that would be a new requirement.
For the Service ↔ NFVO case, Is there an equivalent NFVO/Orchestrator list in the ESR? The esr-nfvo-list would need the same set of info as the VNFM case. If the esr-nfvo-list does exist, I think that there needs to be a reference from the Service Instance record to the specific NFVO instance that deployed the Service. Is there such a reference? If not, I think that would be a new requirement.
Bo Lv can comment more on the current ESR capabilities, but I believe there are only 3 kinds of systems so far: EMS, VNFM and third-party SDNC.
ESR could be extended to handle VNFO as another kind of system.
William LaMont will check for existing migration utility that handles this use case (changing the key from one existing attribute to another)
James Forsyth will socialize the breaking change on the PNF in the next PTL call so clients can prepare to do a search for ?pnf-name=${pnf-name} instead of /pnfs/pnf/${pnf-name}. They also need to handle doing the PUT operation differently - Added to PTL agenda PTL 2019-02-19
if I had some AAI data with attributes that are strings but nominally contain date/timestamps, is there a way to query for a particular range of values?
is there a way to do partial match? regex? PUT /aai/v13/query?format=raw
The schema-service is ready. Currently it provides file-sharing capabilities in terms of schema/edgerule files.
In order for GraphGraph to take advantage of the schema parsing/processing in schema-service additional abstractions have to be implemented on top of the crude file2string functionality currently in schema-service.
Open question:
What schema abstractions are planned in schema-service?
When will the abstractions (from point 1.) be implemented?
Venkata Harish Kajurwill ask Manisha Aggarwalif the current functionality of the schema-service is the final version for Dublin and if there will be further enhancements in next releases.
list of all schema nodes/items (like vserver, tenant, p-interfaces..) for example on a REST path /schemas/{schema}/nodes
all relevant attributes of a given node/item for example on REST path /schemas/{schema}/nodes/{node}
edges/relationships with their attributes between schema nodes/items (for example on REST path /schemas/{schema}/edges where you specify a "from" "to" schema items as query params)
subgraph of the schema, where you specify 1. initial (root) items/node (like tenant or vserver) 2. schema version and 3. number of parent/cousin/child hops from the initial item/node
all paths in a given schema graph between 2 items/nodes (like vserver and tenant) for a given schema version
edges in the schema graph should be composed of edges in the schema file + edges created from the edgerules file
edges should contain basic attributes when delivered via the subgraph call (like parent/child relationship and important properties from edgerules) and have additional (or all) attributes when queries via /schemas/{schema}/edges REST endpoint.
AAF will generate certificates to the be used by the containers at startup; AAI services should use the run-time generated certs instead of the ones that are in the repos or oom charts.
In dublin the services will mount a volume with certificates. This is on the roadmap for Dublin as a feature.
is this for all service and/or HAProxy?
Where are the certificates coming from (OOM/gerrit/generated by AAF)
James Forsythwill ask Jonathan Gatham when the certificate init image is going to be available in ONAP and wether it is documented
what would be the approach to backup an entire ONAP instance particualarly SDC, AAI, SDNC data ? would it be a script with all the references to the helm deploy releases or something that does a helm list and then for each entry does the ark backup ?
What is the AAI strategy for backup and restore?
What is the overall ONAP strategy for backup and restore?
Should it be unified with the data migration strategy as per "Hbase to Cassandra migration" on 2018-11-14 AAI Meeting Notes?
James Forsythwill raise the topic of having backups and restore functionality in ONAP - if it is feasible, on the roadmap and what others PTL think
Jimmy didn't directly raise the topic but there was movement - Keong Lim asked "if istio service mesh is a no-go, is there a replacement for secure onap communications? is backup/restore/upgradability included in s3p?"
hector has discovered that the stress test jar (liveness probe?) in aai-cassandra is hammering the cpu/ram/hd on the vm that aai is on - this breaks the etcd cluster (not the latency/network issues we suspected that may cause pod rescheduling)
Is there something that should be tweaked in AAI config? Or documentation on the recommended setup to run the VM?
I'll come to the next AAI meet (conflicts with pomba meet) -
Jira
server
ONAP JIRA
serverId
425b2b0a-557c-3c0c-b515-579789cceedb
key
LOG-376
20190108 work continues to find the cause - I see 7 vCore spikes on cassandra as well as a saturated logstash on that particular vm - we are no longer a DaemonSet (13 instances on a 13+1 cluster) - I will reduce the current ReplicaSet from 5 to 2 or 1 until I can label the nodes and/or find out what is causing ls to saturate - Prudence Au and Sanjay Agraharam mentioned cassandra - I have seen cs high on several "top" sessions - will post screen caps - bottom line is correlation - I have a 2nd cluster where I can just run aai,dmaap and log
f. AAI team wanted to get notified of AAI Cassandra issues automatically i. Can we setup a Nagios or equivalent to monitor both rancher/k8 and the applications for rancher/k8 issues ?
Keep an eye out for new issues!
Question: Keong Lim should this be part of a larger A&AI monitoring and failure prevention initiative
21
Modeling team R4 discussion, including extra AAI attributes in a model-driven way
9)“vhn-portal-url”?“Bandwidth”,"QoS","SLA",etc, attribtutes that not all the services need but still need to be stored in certain service instance: stored as a schemaless field on the service-instance vertex (Chesla will follow up) (my concerns: according to the call, is that ok if we set a "global-type of service" and a "customized-type of service", then mapped it with internal descriptor, and A&AI's model only stores global type in service instance's schema, but stores the customer-faced attributes of service in a schemaless way? Chesla WechslerKevin ScaggsAndy Mayer)
The service-instance already uses a "metadata" relationship, which can store an arbitrary list of key-value pairs, but perhaps AAI should extend the use of the "properties" element, which is also an arbitrary list of name-value pairs or the "extra-properties" element, which is also an arbitrary list of name-value pairs.
15th Nov: Having seen Chesla's presentation, it should be called "Model-driven schema" rather than "schemaless" behaviour, since the idea is that the changes are controlled by SDC modelling. Seems aligned to the eventual goal in AAI Schema Service Use Case Proposals and AAI Schema Service.
Could we disable unused (i.e. not integrated) A&AI web services, so that the deployment is faster and the resource footprint is smaller? e.g. Champ (any other ws?)
Motivation: Decrease the resource footprint for A&AI (ONAP) deployments
Idea: we could support 2 different deployments 1. full (normal) deployment and 2. barebones deployment. The point of the "barebone" deployment would be to deploy only the essential services necessary for proper functioning of A&AI (leaving out services like cacher, sparky, graphadmin, having 1 cassandra node instead of 3 or 5 etc).
In order to reduce hardware/cloud costs (mainly the memory footprint) it could be beneficial to support a minimalistic A&AI deployment.
1st Nov:
Venkata Harish Kajurand Pavel Paroulek - investigate how to disable/enable charts in A&AI so we can create a core group of pods which handle the use-cases and than extended group will all the services. Consider a group of unused/unintegrated services (like Champ). Consider other possible groups (like GUI?)
James Forsythcreates a JIRA ticket to define the list of AAI subprojects and create the categories (essential, full "experience") for the OOM deployment
Keong Limwill ask the teams about the data mentioned above.
James Forsyth will remind Christopher to create a LF account
James Forsyth will ask for a Fitness test which tests the correctness of conversion from name-query to custom-quesry
James Forsyth will try to find out if any documentation can be provided from ecomp regarding the named-queries (how they work and how to read the named-query files)