summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorFrank Brockners <fbrockne@cisco.com>2016-10-19 13:28:24 -0700
committerFrank Brockners <fbrockne@cisco.com>2016-10-19 13:34:08 -0700
commitbbace7a732cbe9ffbc0786502b11434170bf6ebf (patch)
treee308f50e9a1344bed89c101c94addf7dc990a24d
parenta840f8c6d0647238b30dea8a504eee6bdba64946 (diff)
Scenario documentation for os-odl_l2-fdio-ha
Change-Id: I1b1dec0e95d96809e30393efab268c81317645e2 Signed-off-by: Frank Brockners <fbrockne@cisco.com>
-rwxr-xr-xdocs/scenarios/os-odl_l2-fdio-ha/FDS-L3-tenant-connectivity.pngbin0 -> 50947 bytes
-rwxr-xr-xdocs/scenarios/os-odl_l2-fdio-ha/FDS-basic-callflow.jpgbin0 -> 148454 bytes
-rwxr-xr-xdocs/scenarios/os-odl_l2-fdio-ha/FDS-odl_l2-overview.pngbin0 -> 90635 bytes
-rw-r--r--docs/scenarios/os-odl_l2-fdio-ha/index.rst18
-rwxr-xr-xdocs/scenarios/os-odl_l2-fdio-ha/os-odl_l2-fdio-ha-colorado2_1.pngbin0 -> 174442 bytes
-rwxr-xr-xdocs/scenarios/os-odl_l2-fdio-ha/scenario.description.rst306
6 files changed, 324 insertions, 0 deletions
diff --git a/docs/scenarios/os-odl_l2-fdio-ha/FDS-L3-tenant-connectivity.png b/docs/scenarios/os-odl_l2-fdio-ha/FDS-L3-tenant-connectivity.png
new file mode 100755
index 0000000..9de77e5
--- /dev/null
+++ b/docs/scenarios/os-odl_l2-fdio-ha/FDS-L3-tenant-connectivity.png
Binary files differ
diff --git a/docs/scenarios/os-odl_l2-fdio-ha/FDS-basic-callflow.jpg b/docs/scenarios/os-odl_l2-fdio-ha/FDS-basic-callflow.jpg
new file mode 100755
index 0000000..96464f8
--- /dev/null
+++ b/docs/scenarios/os-odl_l2-fdio-ha/FDS-basic-callflow.jpg
Binary files differ
diff --git a/docs/scenarios/os-odl_l2-fdio-ha/FDS-odl_l2-overview.png b/docs/scenarios/os-odl_l2-fdio-ha/FDS-odl_l2-overview.png
new file mode 100755
index 0000000..2755099
--- /dev/null
+++ b/docs/scenarios/os-odl_l2-fdio-ha/FDS-odl_l2-overview.png
Binary files differ
diff --git a/docs/scenarios/os-odl_l2-fdio-ha/index.rst b/docs/scenarios/os-odl_l2-fdio-ha/index.rst
new file mode 100644
index 0000000..65cf148
--- /dev/null
+++ b/docs/scenarios/os-odl_l2-fdio-ha/index.rst
@@ -0,0 +1,18 @@
+.. OPNFV - Open Platform for Network Function Virtualization
+.. This work is licensed under a Creative Commons Attribution 4.0 International License.
+.. http://creativecommons.org/licenses/by/4.0
+
+
+*********************************************************************
+Fast Data Stacks Scenario: os-odl_l2-fdio-ha Overview and Description
+*********************************************************************
+
+Scenario: "OpenStack - Opendaylight (L2) - FD.io" (apex-os-odl_l2-fdio-ha)
+is a scenario developed as part of the FastDataStacks
+OPNFV project.
+
+.. toctree::
+ :numbered:
+ :maxdepth: 2
+
+ scenario.description.rst
diff --git a/docs/scenarios/os-odl_l2-fdio-ha/os-odl_l2-fdio-ha-colorado2_1.png b/docs/scenarios/os-odl_l2-fdio-ha/os-odl_l2-fdio-ha-colorado2_1.png
new file mode 100755
index 0000000..aa23495
--- /dev/null
+++ b/docs/scenarios/os-odl_l2-fdio-ha/os-odl_l2-fdio-ha-colorado2_1.png
Binary files differ
diff --git a/docs/scenarios/os-odl_l2-fdio-ha/scenario.description.rst b/docs/scenarios/os-odl_l2-fdio-ha/scenario.description.rst
new file mode 100755
index 0000000..d8787c6
--- /dev/null
+++ b/docs/scenarios/os-odl_l2-fdio-ha/scenario.description.rst
@@ -0,0 +1,306 @@
+.. OPNFV - Open Platform for Network Function Virtualization
+.. This work is licensed under a Creative Commons Attribution 4.0 International License.
+.. http://creativecommons.org/licenses/by/4.0
+
+Scenario: "OpenStack - OpenDaylight (Layer 2) - FD.io"
+======================================================
+
+Scenario: apex-os-odl_l2-fdio-ha
+
+"apex-os-odl_l2-fdio-ha" is a scenario developed as part of the
+FastDataStacks OPNFV project. The main components of the
+"apex-os-odl_l2-fdio-ha" scenario are:
+
+ - APEX (TripleO) installer (please also see APEX installer documentation)
+ - Openstack (in HA configuration)
+ - OpenDaylight controller (non-clustered) controlling layer 2 networking
+ - FD.io/VPP virtual forwarder for tenant networking
+
+Introduction
+============
+
+NFV and virtualized high performance applications, such as video processing,
+require a "fast data stack" solution that provides both carrier grade
+forwarding performance, scalability and open extensibility, along with
+functionality for realizing application policies and controlling a complex
+network topology.
+
+A solution stack is only as good as its foundation. Key foundational assets for
+NFV infrastructure are
+ * The virtual forwarder: The virtual forwarder needs to be a feature rich,
+ high performance, highly scale virtual switch-router. It needs to leverage
+ hardware accelerators when available and run in user space.
+ In addition, it should be modular and easily extensible.
+ * Forwarder diversity: A solution stack should support a variety of
+ forwarders, hardware forwarders (physical switches and routers)
+ as well as software forwarders. This way virtual and physical
+ forwarding domains can be seamlessly glued together.
+ * Policy driven connectivity: Connectivity should respect and
+ reflect different business
+
+In order to meet the desired qualities of an NFV infrastructure, the
+following components were chosen for the "Openstack - OpenDaylight
+ - FD.io/VPP" scenario:
+ * FD.io Vector Packet Processor (VPP) - a highly scalable,
+ high performance, extensible virtual forwarder
+ * OpenDaylight Controller - an extensible controller platform which
+ offers the ability to separate business logic from networking
+ constructs, supports a diverse set of network devices
+ (virtual and physical) via the "group based policy (GBP)"
+ component, and can be clustered to achieve a highly available
+ deployment.
+
+The "Openstack - OpenDaylight - FD.io/VPP" scenario provides the capability to
+realize a set of use-cases relevant to the deployment of NFV nodes instantiated
+by means of an Openstack orchestration system on FD.io/VPP enabled compute
+nodes. The role of the Opendaylight network controller in this integration is
+twofold. It provides a network device configuration and topology abstraction
+via the Openstack Neutron interface, while providing the capability to realize
+more complex network policies by means of Group Based Policies. Furthermore it
+also provides the capabilities to monitor as well as visualize the operation of
+the virtual network devices and their topologies.
+In supporting the general use-case of instantiatiting an NFV instance, two
+specific types of network transport use cases are realized:
+
+ * NFV instances with VPP data-plane forwarding using a VLAN provider network
+ * NFV instances with VPP data-plane forwarding using a VXLAN overlay
+ transport network
+
+A deployment of the "apex-os-odl_l2-fdio-ha" scenario consists of 4 or more
+servers:
+
+ * 1 Jumphost hosting the APEX installer - running the Undercloud
+ * 3 Controlhosts, which runs the Overcloud as well as OpenDaylight
+ as a network controller (OpenDaylight only runs on one Controlhost)
+ * 2 or more Computehosts
+
+.. image:: FDS-odl_l2-overview.png
+
+Tenant networking leverages FD.io/VPP. Open VSwitch (OVS) is used for all other
+connectivity, in particular the connectivity to public networking / the
+Internet (i.e. br-ext) is performed via OVS as in any standard OpenStack
+deployment. The OpenDaylight network controller is used to setup and manage
+layer 2 networking for the scenario. Tenant networking can either leverage
+VXLAN (in which case a full mesh of VXLAN tunnels is created) or VLANs. Layer 3
+connectivity for a tenant network is provided centrally via qrouter on the
+control node. As in a standard OpenStack deployment, the Layer3 agent
+configures the qrouter and associated rulesets for security (security groups)
+and NAT (floating IPs). Public IP network connectivity for a tenant network is
+provided by interconnecting the VPP-based bridge domain representing the tenant
+network to qrouter using a tap interface. The setup is depicted below:
+
+.. image:: FDS-L3-tenant-connectivity.png
+
+With high availability factored in the setup looks like the following.
+
+.. image:: os-odl_l2-fdio-ha-colorado2_1.png
+
+Note that the picture only shows two Controllernodes for reasons of
+simplicity. A HA deployment will always include 3 Controllernodes.
+
+
+Features of the scenario
+------------------------
+
+Main features of the "apex-os-odl_l2-fdio-ha" scenario:
+
+ * Automated installation using the APEX installer
+ * Fast and scalable tenant networking using FD.io/VPP as forwarder
+ * Layer 2 networking using VLANs or VXLAN, managed and
+ controlled through OpenDaylight
+ * Layer 3 connectivitiy for tenant networks supplied centrally on
+ the Control node through standard OpenStack mechanisms.
+ All layer 3 features apply, including floating IPs (i.e. NAT)
+ and security groups.
+ * Manual and automatic (via DHCP) addressing on tenant networks
+ * OpenStack high availability
+
+Scenario components and composition
+===================================
+
+The apex-os-odl_l2-fdio-ha scenario combines components from three key open
+source projects: OpenStack, OpenDaylight, and Fast Data (FD.io). The key
+components that realize the apex-os-odl_l2-fdio-ha scenario and which differ
+from a regular, OVS-based scenario, are the OpenStack ML2 OpenDaylight plugin,
+OpenDaylight Neutron Northbound, OpenDaylight Group Based Policy, OpenDaylight
+Virtual Bridge Domain Manager, FD.io Honeycomb management agent and FD.io
+Vector Packet Processor (VPP).
+
+Here's a more detailed list of the individual software components involved:
+
+**Openstack Neutron ML2 ODL Plugin**: Handles Neutron data base synchronization
+and interaction with the southbound Openstack controller using HTTP.
+
+**OpenDaylight Neutron Nothbound & Neutron MD-SAL Entry Store**: Presents a
+Neutron (v2) extended HTTP API servlet for interaction with Openstack Neutron.
+It validates and stores the received Neutron data in the MD-SAL data store
+against the Neutron yang model driven.
+
+**OpenDaylight Neutron Mapper**: The Neutron Mapper listens to Neutron data
+change events and is responsible for using Neutron data in creating Group Based
+Policy Data objects, e.g. GBP End-Points, Flood-Domains. A GBP End Point
+represents a specific NFV/VM port and its identity as derived from a Neutron
+Port. The mapped data is stored using the GBP End Point yang model and an
+association between the GBP End-Point and its Neutron object is maintained in
+the Neutron-GBP map.
+
+**OpenDaylight Group Based Policy (GBP) Entities store**: Stores for the GBP
+data artifacts against the GBP YANG schemas.
+
+**Neutron Group Based Policy Map store**: Stores the bi-lateral relation
+between an End-Point and its corresponding Neutron object. Neutron-GBP map;
+keyed by Neutron object type, port, and Neutron UUID, gives the GBP End-Point,
+Flood domain respectively. GBP-Neutron map keyed by GBP object type, end-point.
+
+**Neutron VPP Renderer Mapper**: The Neutron VPP Renderer Mapper listens to
+Neutron Store data change events, as well as being able to access directly the
+store, and is responsible for converting Neutron data specifically required to
+render a VPP node configuration with a given End Point, e.g. the virtual host
+interface name assigned to a vhostuser socket.. The mapped data is stored in
+the VPP info data store.
+
+**VPP Info Store**: Stores VPP specific information regarding End-Points, Flood
+domains with VLAN, etc.
+
+**GBP Renderer Manager**: The GBP Renderer Manager is the central point for
+dispatching of data to specific device renderers. It uses the information
+derived from the GBP end-point and its topology entries to dispatch the task of
+configuration to a specific device renderer by writing a renderer policy
+configuration into the registered renderer's policy store. The renderer manager
+also monitors, by being a data change listener on the VPP Renderer Policy
+States, for any errors in the application of a rendered configuration.
+
+**Renderer Policy Config Store**: The store's schema serves as the API between
+the Renderer Manager and specific Renderers like the VPP Renderer. The store
+uses a a YANG modeled schema to represent all end-point and associated GBP
+policy data.
+
+**Topology Entries Store**: The yang model based MD-SAL topology store serves
+two fundamental roles: 1. It maintains a topological representation of the GBP
+End Points, in the context of customer networks. 2. It maintains an association
+of each (VPP) compute node's physical interfaces to their neutron provider
+network (e.g. The association between an ethernet interface and a Neutron
+provider network).
+
+**VPP Renderer**: The VPP Renderer registers an instance for VPP nodes with the
+Renderer Manager by means of inserting operational data into the Renderer
+Policy config store. It acts as a listener on the Renderer Policy consumes via
+the GBP Policy API data + the specific VPP End Point data, to drive the
+configuration of VPP devices using NETCONF Services.
+More specifically, the renderer generates:
+
+ * vhost user port configuration that corresponds to the VM port configuration
+ * VPP bridge instances corresponding to the GBP flood domain
+ * port or traffic filtering configuration, in accordance with the GBP policy.
+
+The VPP Renderer also interacts with the Virtual Bridge Domain Service, by
+means of the VBD store, in order to establish connectivity between VPP nodes in
+a bridge domain. For this it uses the VPP device name, and the flood domain
+data derived from the VPP Info and End-Point data respectively. For the
+executed configuration operations it updates state in the Renderer policy state
+store.
+
+**Virtual Bridge Domain (VBD) Store and Manager**: The virtual bridge domain
+manager is responsible for configuring the VxLAN overlay tunnel infrastructure
+to arrive at a desired bridged topology between multiple (VPP) compute nodes.
+VDB configures VXLAN tunnels always into a full-mesh with split-horizon group
+forwarding applied on any domain facing tunnel interface (i.e. forwarding
+behavior will be that used for VPLS).
+
+**NETCONF Mount Point Service & Connector**: Collectively referred to as
+Netconf Services, provide the NETCONF interface for accessing VPP configuration
+and operational data stores that are represented as NETCONF mounts.
+
+**Virtual Packet Processor (VPP) and Honeycomb server**: The VPP is the
+accelerated data plane forwarding engine relying on vhost user interfaces
+towards Virtual Machines created by the Nova Agent. The Honeycomb NETCONF
+configuration server is responsible for driving the configuration of the VPP,
+and collecting the operational data.
+
+**Rendered Policy State Store**: Stores data regarding the execution of
+operations performed by a given renderer.
+
+**Nova Agent**: The Nova Agent, a sub-component of the overall Openstack
+architecture, is responsible for interacting with the compute node's host
+Libvirt API to drive the life-cycle of Virtual Machines. It, along with the
+compute node software, are assumed to be capable of supporting vhost user
+interfaces.
+
+The picture below show a basic end to end call flow for creating a Neutron
+vhostuser port on VPP using a GBP renderer. It showcases how the different
+component described above interact.
+
+.. image:: FDS-basic-callflow.jpg
+
+Scenario Configuration
+======================
+
+To enable the "apex-os-odl_l2-fdio-ha" scenario check the appropriate
+settings in the APEX configuration files. Those are typically found in
+/etc/opnfv-apex.
+
+File "deploy_settings.yaml" choose opendaylight as controller with version
+"boron" and enable vpp as forwarder. "hugepages" need to set to a
+sufficiently large value for VPP to work. The default value for VPP is
+1024, but this only allows for a few VMs to be started. If feasible,
+choose a significantly larger number on the compute nodes::
+
+ global_params:
+ ha_enabled: true
+
+ deploy_options:
+ sdn_controller: opendaylight
+ sdn_l3: false
+ odl_version: boron
+ tacker: true
+ congress: true
+ sfc: false
+ vpn: false
+ vpp: true
+ dataplane: fdio
+ performance:
+ Controller:
+ kernel:
+ hugepages: 1024
+ hugepagesz: 2M
+ intel_iommu: 'on'
+ iommu: pt
+ Compute:
+ kernel:
+ hugepagesz: 2M
+ hugepages: 2048
+ intel_iommu: 'on'
+ iommu: pt
+
+
+Validated deployment environments
+=================================
+
+The "os-odl_l2-fdio-ha" scenario has been deployed and tested
+on the following sets of hardware:
+ * Linux Foundation lab (Chassis: Cisco UCS-B-5108 blade server,
+ NICs: 8 external / 32 internal 10GE ports,
+ RAM: 32G (4 x 8GB DDR4-2133-MHz RDIMM/PC4-17000/single rank/x4/1.2v),
+ CPU: 3.50 GHz E5-2637 v3/135W 4C/15MB Cache/DDR4 2133MHz
+ Disk: 1.2 TB 6G SAS 10K rpm SFF HDD) see also:
+ https://wiki.opnfv.org/display/pharos/Lflab+Hosting
+ * OPNFV CENGN lab (https://wiki.opnfv.org/display/pharos/CENGN+Pharos+Lab)
+ * Cisco internal development labs (UCS-B and UCS-C)
+
+Limitations, Issues and Workarounds
+===================================
+
+There are no known issues. Note that only OpenStack is deployed in
+HA mode. OpenDaylight clustering is expected to be added in a future
+revision of this scenario.
+
+References
+==========
+
+
+ * FastDataStacks OPNFV project wiki: https://wiki.opnfv.org/display/fds
+ * Fast Data (FD.io): https://fd.io/
+ * FD.io Vector Packet Processor (VPP): https://wiki.fd.io/view/VPP
+ * OpenDaylight Controller: https://www.opendaylight.org/
+ * OPNFV Colorado release - more information: http://www.opnfv.org/colorado
+