diff options
author | Ryota Mibu <r-mibu@cq.jp.nec.com> | 2017-02-17 04:36:05 +0000 |
---|---|---|
committer | Gerrit Code Review <gerrit@opnfv.org> | 2017-02-17 04:36:05 +0000 |
commit | f3ab498aaddb27f6f598a84e2dbe0203ced6d666 (patch) | |
tree | fc7b2be2681db87adc1eb935e6fdcc93a8bc1645 /docs/development/requirements/02-use_cases.rst | |
parent | 5d9c24fd28bcc02243306a8c96d0c68809523343 (diff) | |
parent | d0b22e1d856cf8f78e152dfb6c150e001e03dd52 (diff) |
Merge "Update docs structure according to new guidelines in https://wiki.opnfv.org/display/DOC"
Diffstat (limited to 'docs/development/requirements/02-use_cases.rst')
-rw-r--r-- | docs/development/requirements/02-use_cases.rst | 195 |
1 files changed, 195 insertions, 0 deletions
diff --git a/docs/development/requirements/02-use_cases.rst b/docs/development/requirements/02-use_cases.rst new file mode 100644 index 00000000..0a1f6413 --- /dev/null +++ b/docs/development/requirements/02-use_cases.rst @@ -0,0 +1,195 @@ +.. This work is licensed under a Creative Commons Attribution 4.0 International License. +.. http://creativecommons.org/licenses/by/4.0 + +Use cases and scenarios +======================= + +Telecom services often have very high requirements on service performance. As a +consequence they often utilize redundancy and high availability (HA) mechanisms +for both the service and the platform. The HA support may be built-in or +provided by the platform. In any case, the HA support typically has a very fast +detection and reaction time to minimize service impact. The main changes +proposed in this document are about making a clear distinction between fault +management and recovery a) within the VIM/NFVI and b) High Availability support +for VNFs on the other, claiming that HA support within a VNF or as a service +from the platform is outside the scope of Doctor and is discussed in the High +Availability for OPNFV project. Doctor should focus on detecting and remediating +faults in the NFVI. This will ensure that applications come back to a fully +redundant configuration faster than before. + +As an example, Telecom services can come with an Active-Standby (ACT-STBY) +configuration which is a (1+1) redundancy scheme. ACT and STBY nodes (aka +Physical Network Function (PNF) in ETSI NFV terminology) are in a hot standby +configuration. If an ACT node is unable to function properly due to fault or any +other reason, the STBY node is instantly made ACT, and affected services can be +provided without any service interruption. + +The ACT-STBY configuration needs to be maintained. This means, when a STBY node +is made ACT, either the previously ACT node, after recovery, shall be made STBY, +or, a new STBY node needs to be configured. The actual operations to +instantiate/configure a new STBY are similar to instantiating a new VNF and +therefore are outside the scope of this project. + +The NFVI fault management and maintenance requirements aim at providing fast +failure detection of physical and virtualized resources and remediation of the +virtualized resources provided to Consumers according to their predefined +request to enable applications to recover to a fully redundant mode of +operation. + +1. Fault management/recovery using ACT-STBY configuration (Triggered by critical + error) +2. Preventive actions based on fault prediction (Preventing service stop by + handling warnings) +3. VM Retirement (Managing service during NFVI maintenance, i.e. H/W, + Hypervisor, Host OS, maintenance) + +Faults +------ + +.. _uc-fault1: + +Fault management using ACT-STBY configuration +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + +In :numref:`figure1`, a system-wide view of relevant functional blocks is +presented. OpenStack is considered as the VIM implementation (aka Controller) +which has interfaces with the NFVI and the Consumers. The VNF implementation is +represented as different virtual resources marked by different colors. Consumers +(VNFM or NFVO in ETSI NFV terminology) own/manage the respective virtual +resources (VMs in this example) shown with the same colors. + +The first requirement in this use case is that the Controller needs to detect +faults in the NFVI ("1. Fault Notification" in :numref:`figure1`) affecting +the proper functioning of the virtual resources (labelled as VM-x) running on +top of it. It should be possible to configure which relevant fault items should +be detected. The VIM (e.g. OpenStack) itself could be extended to detect such +faults. Alternatively, a third party fault monitoring tool could be used which +then informs the VIM about such faults; this third party fault monitoring +element can be considered as a component of VIM from an architectural point of +view. + +Once such fault is detected, the VIM shall find out which virtual resources are +affected by this fault. In the example in :numref:`figure1`, VM-4 is +affected by a fault in the Hardware Server-3. Such mapping shall be maintained +in the VIM, depicted as the "Server-VM info" table inside the VIM. + +Once the VIM has identified which virtual resources are affected by the fault, +it needs to find out who is the Consumer (i.e. the owner/manager) of the +affected virtual resources (Step 2). In the example shown in :numref:`figure1`, +the VIM knows that for the red VM-4, the manager is the red Consumer +through an Ownership info table. The VIM then notifies (Step 3 "Fault +Notification") the red Consumer about this fault, preferably with sufficient +abstraction rather than detailed physical fault information. + +.. figure:: images/figure1.png + :name: figure1 + :width: 100% + + Fault management/recovery use case + +The Consumer then switches to STBY configuration by switching the STBY node to +ACT state (Step 4). It further initiates a process to instantiate/configure a +new STBY. However, switching to STBY mode and creating a new STBY machine is a +VNFM/NFVO level operation and therefore outside the scope of this project. +Doctor project does not create interfaces for such VNFM level configuration +operations. Yet, since the total failover time of a consumer service depends on +both the delay of such processes as well as the reaction time of Doctor +components, minimizing Doctor's reaction time is a necessary basic ingredient to +fast failover times in general. + +Once the Consumer has switched to STBY configuration, it notifies (Step 5 +"Instruction" in :numref:`figure1`) the VIM. The VIM can then take +necessary (e.g. pre-determined by the involved network operator) actions on how +to clean up the fault affected VMs (Step 6 "Execute Instruction"). + +The key issue in this use case is that a VIM (OpenStack in this context) shall +not take a standalone fault recovery action (e.g. migration of the affected VMs) +before the ACT-STBY switching is complete, as that might violate the ACT-STBY +configuration and render the node out of service. + +As an extension of the 1+1 ACT-STBY resilience pattern, a STBY instance can act as +backup to N ACT nodes (N+1). In this case, the basic information flow remains +the same, i.e., the consumer is informed of a failure in order to activate the +STBY node. However, in this case it might be useful for the failure notification +to cover a number of failed instances due to the same fault (e.g., more than one +instance might be affected by a switch failure). The reaction of the consumer +might depend on whether only one active instance has failed (similar to the +ACT-STBY case), or if more active instances are needed as well. + +Preventive actions based on fault prediction +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + +The fault management scenario explained in :ref:`uc-fault1` can also be +performed based on fault prediction. In such cases, in VIM, there is an +intelligent fault prediction module which, based on its NFVI monitoring +information, can predict an imminent fault in the elements of NFVI. +A simple example is raising temperature of a Hardware Server which might +trigger a pre-emptive recovery action. The requirements of such fault +prediction in the VIM are investigated in the OPNFV project "Data Collection +for Failure Prediction" [PRED]_. + +This use case is very similar to :ref:`uc-fault1`. Instead of a fault +detection (Step 1 "Fault Notification in" :numref:`figure1`), the trigger +comes from a fault prediction module in the VIM, or from a third party module +which notifies the VIM about an imminent fault. From Step 2~5, the work flow is +the same as in the "Fault management using ACT-STBY configuration" use case, +except in this case, the Consumer of a VM/VNF switches to STBY configuration +based on a predicted fault, rather than an occurred fault. + +NFVI Maintenance +---------------- + +VM Retirement +^^^^^^^^^^^^^ + +All network operators perform maintenance of their network infrastructure, both +regularly and irregularly. Besides the hardware, virtualization is expected to +increase the number of elements subject to such maintenance as NFVI holds new +elements like the hypervisor and host OS. Maintenance of a particular resource +element e.g. hardware, hypervisor etc. may render a particular server hardware +unusable until the maintenance procedure is complete. + +However, the Consumer of VMs needs to know that such resources will be +unavailable because of NFVI maintenance. The following use case is again to +ensure that the ACT-STBY configuration is not violated. A stand-alone action +(e.g. live migration) from VIM/OpenStack to empty a physical machine so that +consequent maintenance procedure could be performed may not only violate the +ACT-STBY configuration, but also have impact on real-time processing scenarios +where dedicated resources to virtual resources (e.g. VMs) are necessary and a +pause in operation (e.g. vCPU) is not allowed. The Consumer is in a position to +safely perform the switch between ACT and STBY nodes, or switch to an +alternative VNF forwarding graph so the hardware servers hosting the ACT nodes +can be emptied for the upcoming maintenance operation. Once the target hardware +servers are emptied (i.e. no virtual resources are running on top), the VIM can +mark them with an appropriate flag (i.e. "maintenance" state) such that these +servers are not considered for hosting of virtual machines until the maintenance +flag is cleared (i.e. nodes are back in "normal" status). + +A high-level view of the maintenance procedure is presented in :numref:`figure2`. +VIM/OpenStack, through its northbound interface, receives a maintenance notification +(Step 1 "Maintenance Request") from the Administrator (e.g. a network operator) +including information about which hardware is subject to maintenance. +Maintenance operations include replacement/upgrade of hardware, +update/upgrade of the hypervisor/host OS, etc. + +The consequent steps to enable the Consumer to perform ACT-STBY switching are +very similar to the fault management scenario. From VIM/OpenStack's internal +database, it finds out which virtual resources (VM-x) are running on those +particular Hardware Servers and who are the managers of those virtual resources +(Step 2). The VIM then informs the respective Consumer (VNFMs or NFVO) in Step 3 +"Maintenance Notification". Based on this, the Consumer takes necessary actions +(Step 4, e.g. switch to STBY configuration or switch VNF forwarding graphs) and +then notifies (Step 5 "Instruction") the VIM. Upon receiving such notification, +the VIM takes necessary actions (Step 6 "Execute Instruction" to empty the +Hardware Servers so that consequent maintenance operations could be performed. +Due to the similarity for Steps 2~6, the maintenance procedure and the fault +management procedure are investigated in the same project. + +.. figure:: images/figure2.png + :name: figure2 + :width: 100% + + Maintenance use case + +.. + vim: set tabstop=4 expandtab textwidth=80: |