From 08274b14ccc7531208693a9a09d8c9ab3bdf8b3b Mon Sep 17 00:00:00 2001 From: fuqiao Date: Thu, 30 Mar 2017 10:13:21 +0800 Subject: Modify format issues Modify format issues JIRA: HA-28 Change-Id: I5ab2bee17801d3b0741b83497391935f28602057 Signed-off-by: fuqiao@chinamobile.com Conflicts: docs/development/overview/OPNFV_HA_Guest_APIs-Overview_HLD.rst docs/development/overview/OPNFV_HA_Guest_APIs-Overview_HLD.rst.bak docs/development/overview/index.rst Change-ID: I5ab2bee17801d3b0741b83497391935f28602057 Signed-off-by: fuqiao@chinamobile.com --- .../overview/OPNFV_HA_Guest_APIs-Overview_HLD.rst | 280 ++++++++++++++++++++- docs/development/overview/index.rst | 10 +- 2 files changed, 287 insertions(+), 3 deletions(-) (limited to 'docs/development/overview') diff --git a/docs/development/overview/OPNFV_HA_Guest_APIs-Overview_HLD.rst b/docs/development/overview/OPNFV_HA_Guest_APIs-Overview_HLD.rst index 92ff964..5be0b28 100644 --- a/docs/development/overview/OPNFV_HA_Guest_APIs-Overview_HLD.rst +++ b/docs/development/overview/OPNFV_HA_Guest_APIs-Overview_HLD.rst @@ -1,6 +1,284 @@ +<<<<<<< HEAD .. This work is licensed under a Creative Commons Attribution 4.0 International License. .. http://creativecommons.org/licenses/by/4.0 .. (c) ==================================== OPNFV HA Guest APIs -- Overview -==================================== \ No newline at end of file +==================================== +======= +Overview +===================================================================== + +:author: Greg Waines +:organization: Wind River Systems +:organization: OPNFV - High Availability +:status: Draft - PROPOSAL +:date: March 2017 +:revision: 1.5 + +:abstract: This document presents a PROPOSAL for a set of new + optional capabilities where the OpenStack Cloud messages + into the Guest VMs in order to provide improved Availability + of the hosted VMs. The initial set of new capabilities + include: enabling the detection of and recovery from internal + VM faults and providing a simple out-of-band messaging service + to prevent scenarios such as split brain. + + +.. sectnum:: + +.. contents:: Table of Contents + + + +Introduction +===================================================================== + + This document provides an overview and rationale for a PROPOSAL + of a set of new capabilities where the OpenStack Cloud messages + into the Guest VMs in order to provide improved Availability + of the hosted VMs. + + The initial set of new capabilities specifically include: + + - VM Heartbeating and Health Checking + - VM Peer State Notification and Messaging + + All of these capabilities leverage Host-to-Guest Messaging + Interfaces / APIs which are built on a messaging service between the + OpenStack Host and the Guest VM that uses a simple low-bandwidth + datagram messaging capability in the hypervisor and therefore has no + requirements on OpenStack Networking, and is available very early + after spawning the VM. + + For each capability, the document outlines the interaction with + the Guest VM, any key technologies involved, the integration into + the larger OpenStack and OPNFV Architectures (e.g. interactions + with VNFM), specific OPNFV HA Team deliverables, and the use cases + for how availability of the hosted VM is improved. + + The intent is for the OPNFV HA Team to review the proposals of this + document with the related other teams in OPNFV (Doctor and Management + & Orchestration (MANO)) and OpenStack (Nova). + + + + +Messaging Layer +======================================================================== + + The Host-to-Guest messaging APIs used by the services discussed + in this document use a JSON-formatted application messaging layer + on top of a ‘virtio serial device’ between QEMU on the OpenStack Host + and the Guest VM. JSON formatting provides a simple, humanly readable + messaging format which can be easily parsed and formatted using any + high level programming language being used in the Guest VM (e.g. C/C++, + Python, Java, etc.). Use of the ‘virtio serial device’ provides a + simple, direct communication channel between host and guest which is + independent of the Guest’s L2/L3 networking. + + The upper layer JSON messaging format is actually structured as a + hierarchical JSON format containing a Base JSON Message Layer and an + Application JSON Message Layer: + + - the Base Layer provides the ability to multiplex different groups + of message types on top of a single ‘virtio serial device’ + e.g. + + + heartbeating and healthchecks, + + server group messaging, + + and + + - the Application Layer provides the specific message types and + fields of a particular group of message types. + + + +VM Heartbeating and Health Checking +============================================================================ + + Normally OpenStack monitoring of the health of a Guest VM is limited + to a black-box approach of simply monitoring the presence of the + QEMU/KVM PID containing the VM. + + VM Heartbeating and Health Checking provides a heartbeat service to monitor + the health of guest application(s) within a VM running under the OpenStack + Cloud. Loss of heartbeat or a failed health check status will result in a + fault event being reported to OPNFV's DOCTOR infrastructure for alarm + identification, impact analysis and reporting. This would then enable VNF + Managers (VNFMs) listening to OPNFV's DOCTOR External Alarm Reporting through + Telemetry's AODH, to initiate any required fault recovery actions. + + .. image:: OPNFV_HA_Guest_APIs-Overview_HLD-Guest_Heartbeat-FIGURE-1.png + + The VM Heartbeating and Health Checking functionality is enabled on + a VM through a new flavor extraspec indicating that the VM supports + and wants to enable Guest Heartbeating. An extension to Nova Compute uses + this extraspec to setup the required 'virtio serial device' for Host-to-Guest + messaging, on the QEMU/KVM instance created for the VM. + + A daemon within the Guest VM will register with the OpenStack Guest + Heartbeat Service on the compute node to initiate the heartbeating on itself + (i.e. the Guest VM). The OpenStack Compute Node will start heartbeating the + Guest VM, and if the heartbeat fails, the OpenStack Compute Node will report + the VM Fault thru DOCTOR and ultimately VNFM will see this thru NOVA VM + State Chagne Notifications thru AODH. I.e. VNFM wouild see the VM Heartbeat + Failure events in teh same way it sees all other VM Faults, thru DOCTOR + initiated VM state changes. + + Part of the Guest VM's registration process is the specification of the + heartbeat interval in msecs. I.e. the registering Guest VM specifies the + heartbeating interval. + + Guest heartbeat works on a challenge response model. The OpenStack + Guest Heartbeat Service on the compute node will challenge the registered + Guest VM daemon with a message each interval. The registered Guest VM daemon + must respond prior to the next interval with a message indicating good health. + If the OpenStack Host does not receive a valid response, or if the response + specifies that the VM is in ill health, then a fault event for the Guest VM + is reported to the OpenStack Guest Heartbeat Service on the controller node which + will report the event to OPNFV's DOCTOR (i.e. thru the OpenStack Vitrage data + source APIs). + + The registered Guest VM daemon's response to the challenge can be as simple + as just immediately responding with OK. This alone allows for detection of + a failed or hung QEMU/KVM instance, or a failure of the OS within the VM to + schedule the registered Guest VM's daemon or failure to route basic IO within + the Guest VM. + + However the registered Guest VM daemon's response to the challenge can be more + complex, running anything from a quick simple sanity check of the health of + applications running in the Guest VM, to a more thorough audit of the + application state and data. In either case returning the status of the + health check enables the OpenStack host to detect and report the event in order + to initiate recovery from application level errors or failures within the Guest VM. + + In summary, the deliverables of this activity would be: + + - Host Deliverables: (OpenStack and OPNFV blueprints and implementation) + + + an OpenStack Nova or libvirt extension to interpret the new flavor extraspec and + if present setup the required 'virtio serial device' for Host-to-Guest + heartbeat / health-check messaging, on the QEMU/KVM instance created + for the VM, + + an OPNFV Base Host-to-Guest Msging Layer Agent for multiplexing of Application + Layer messaging over the 'virtio serial device' to the VM, + + an OPNFV Heartbeat / Health-Check Compute Agent for local heartbeating of VM + and reporting of failures to the OpenStack Controller, + + an OPNFV Heartbeat / Health-check Server on the OpenStack Controller for + receiving VM failure notifications and reporting these to Vitrage thru + Vitrage's Data Source API, + + - Guest Deliverables: + + + a Heartbeat / Health-Check Message Specification covering + + - Heartbeat / Health-Check Application Layer JSON Protocol, + - Base Host-to-Guest JSON Protocol, + - Details on the use of the underlying 'virtio serial device', + + + a Reference Implementation of the Guest-side support of + Heartbeat / Health-check containing the peer protocol layers + within the Guest. + + - will provide code and compile instructions, + - Guest will compile based on its specific OS. + + This proposal requires review with OPNFV's Doctor and Management & Orchestration + teams, and OpenStack's Nova Team. + + + +VM Peer State Notification and Messaging +=================================================================================== + + Server Group State Notification and Messaging is a service to provide + simple low-bandwidth datagram messaging and notifications for servers that + are part of the same server group. This messaging channel is available + regardless of whether IP networking is functional within the server, and + it requires no knowledge within the server about the other members of the group. + + NOTE: A Server Group here is the OpenStack Nova Server Group concept where VMs + are grouped together for purposes of scheduling. E.g. A specific Server Group + instance can specify whether the VMs within the group should be scheduled to + run on the same compute host or different compute hosts. A 'peer' VM in the + context of this section refers to a VM within the same Nova Server Group. + + This Server Group Messaging service provides three types of messaging: + + - Broadcast: this allows a server to send a datagram (size of up to 3050 bytes) + to all other servers within the server group. + - Notification: this provides servers with information about changes to the + (Nova) state of other servers within the server group. + - Status: this allows a server to query the current (Nova) state of all servers within + the server group (including itself). + + A Server Group Messaging entity on both the controller node and the compute nodes + manage the routing of of VM-to-VM messages through the platform, leveraging Nova + to determine Server Group membership and compute node locations of VMs. The Server + Group Messaging entity on the controller also listens to Nova VM state change notifications + and querys VM state data from Nova, in order to provide the VM query and notification + functionality of this service. + + .. image:: OPNFV_HA_Guest_APIs-Overview_HLD-Peer_Messaging-FIGURE-2.png + + This service is not intended for high bandwidth or low-latency operations. It + is best-effort, not reliable. Applications should do end-to-end acks and + retries if they care about reliability. + + This service provides building block type capabilities for the Guest VMs that + contribute to higher availability of the VMs in the Guest VM Server Group. Notifications + of VM Status changes potentially provide a faster and more accurate notification + of failed peer VMs than traditional peer VM monitoring over Tenant Networks. While + the Broadcast Messaging mechanism provides an out-of-band messaging mechanism to + monitor and control a peer VM under fault conditions; e.g. providing the ability to + avoid potential split brain scenarios between 1:1 VMs when faults in Tenant + Networking occur. + + In summary, the deliverables for Server Group Messaging would be: + + - Host Deliverables: + + + a Nova or libvirt extension to interpret the new flavor extraspec and + if present setup the required 'virtio serial device' for Host-to-Guest + Server Group Messaging, on the QEMU/KVM instance created + for the VM, + + [ leveraging the Base Host-to-Guest Msging Layer Agent from previous section ], + + a Server Group Messaging Compute Agent for implementing the Application Layer + Server Group Messaging JSON Protocol with the VM, and forwarding the + messages to/from the Server Group Messaging Server on the Controller, + + a Server Group Messaging Server on the Controller for routing broadcast + messages to the proper Computes and VMs, as well as listening for Nova + VM State Change Notifications and forwarding these to applicable Computes + and VMs, + + - Guest Deliverables: + + + a Server Group Messaging Message Specification covering + + - Server Group Messaging Application Layer JSON Protocol, + - [ leveraging Base Host-to-Guest JSON Protocol from previous section ], + - [ leveraging Details on the use of the underlying 'virtio serial device' from previous section ], + + + a Reference Implementation of the Guest-side support of + Server Group Messaging containing the peer protocol layers + and Guest Application hooks within the Guest. + + This proposal requires review with OPNFV's Management & Orchestration team and + OpenStack's Nova Team. + + +Conclusion +====================================================================================== + + The PROPOSAL of Reach-thru Guest Monitoring and Services described in this document + leverage Host-to-Guest messaging to provide a number of extended capabilities + that improve the Availability of the hosted VMs. These new capabilities + enable detection of and recovery from internal VM faults and provides a simple + out-of-band messaging service to prevent scenarios such as split brain. + + The integration of these proposed new capabilities into the larger OpenStack and OPNFV + Architectures need to be reviewed with the other related teams in OPNFV (Doctor and + Management & Orchestration (MANO)) and OpenStack (Nova). +>>>>>>> e83e826... Modify format issues diff --git a/docs/development/overview/index.rst b/docs/development/overview/index.rst index 71bdbf4..97fc100 100644 --- a/docs/development/overview/index.rst +++ b/docs/development/overview/index.rst @@ -2,12 +2,18 @@ .. http://creativecommons.org/licenses/by/4.0 .. (c) +<<<<<<< HEAD ********************************* OPNFV HA Guest APIs -- Overview ********************************* +======= + +********************************************************************* +PROPOSAL: Reach-thru Guest Monitoring and Services for High Availability +********************************************************************* +>>>>>>> e83e826... Modify format issues .. toctree:: - :numbered: - :maxdepth: 2 + :maxdepth: 4 OPNFV_HA_Guest_APIs-Overview_HLD.rst -- cgit 1.2.3-korg