Multi-Site Management in OpenStack

Managing multiple Openstack clouds as a single resource pool

In this series of posts, we would be diving into "Tricircle" 
- an open source project that promises to give a "single pane of glass" management over multiple OpenStack environments, using a cascading approach.

As more and more companies are deploying OpenStack, it is becoming clear that there is a need to be able to manage multiple cloud installations. The reasons range from application lifecycle management, through spill-over scenarios and all the way to multi-site Disaster Recovery orchestration.

So why would one care to deploy the same service over multiple environments?

There are multiple reasons, and here are a few:

• Service continuity and geo-redundancy (in case of one site going awry)
• Geo-based load balancing and service locality (in case traffic comes from various places in the globe, or if there are strict quality/latency requirements, or if there are regulatory constraints, etc)
• Cost optimization (in case some resources are cheaper in another place)
• Growth (in case one environment cannot grow enough)
• Resource Utilization (in case you have multiple sites and want to aggregate their resources for better value or easier management)
• Single Configuration (instead of continually synchronizing multiple sub-instances of the service)
• ... (feel free to share additional incentives in the comments).

OpenStack Tricircle

Managing multiple OpenStack instances could be done in several ways, for example by introducing multi-site awareness into each project in OpenStack (which we ruled out due to complexity of evolving all OpenStack projects to do it).

The approach we took in Tricircle was to add a "Top" management OpenStack instance over multiple "Bottom" OpenStack instances.

The "Top" introduces a cascading service layer to delegate APIs downwards, and injects itself into several OpenStack components.

So, how does it feel to use such a "Top" OpenStack instance?

Well, first of all let's define the different users:

1. The Multi-site Tenant Admin (the "User") - Uses the multi-site OpenStack cloud (create VMs, networks, etc.)
2. The Multi-site Admin (the "Admin") - This user can add new sites, and needs to have the necessary credentials on them to put it together

User

For the "User", it is pretty straightforward: when you launch a VM, you get to choose from a list of Data Centers (a new drop box in Horizon), and then from a list of Availability Zones based on your Data Center selection, and that's basically it.

Admin


For the "Admin", you get a new "Add Site" API (in CLI only, at this point), and you need to have substantial knowledge about the "Bottom" sites you are adding (credentials and network-related information which we will cover in the next post).

Some High-Level Architecture

The "Top" Instance

The design principle we took was to reuse OpenStack components in the TOP and bottom layer as much as possible and to mange any OpenStack
deployment without any additional requirements (OpenStack API compatible).

For the top layer we used a non-modified OpenStack API layer to intercept operational requests and handle them in the cascading service.

Doing this required integrating with the different OpenStack core components:

Neutron 

• We introduced a custom core plugin
• Updates are written to the database 
• Operational requests are forwarded to our OpenStack Adaptor service
• Reads and Reports are served directly from the database

Nova

• We implemented a custom Nova Scheduler that runs inside our OpenStack Adaptor service
• We created a Compute Node emulation that runs in the Adaptor service and listens to the Nova Compute service queues
• Our current working assumption is to map "Bottom" sites as "Compute Nodes" that reside on different logical AZs
• The Compute Node emulation instance for each "Bottom" site also aggregates information and statistics that represent the site
• At some point, we plan to let the admin decide how to expose the "Bottom" sites, e.g. different AZs on the "Bottom" site, or all the actual "Compute Nodes", etc. This will create a complete decoupling between the Adaptor and the "Cascading" service.

The "Bottom" Instances

We assume that the "Bottom" sites are unmodified and potentially heterogeneous (in terms of network, configuration and version).

At this stage, our design assumes a centralized Keystone service running on the "Top" (we are planing "Federated Keystone" in the future).

In order to add a "Bottom" site to the multi-site environment, the admin needs to deploy a "Cascaded" service, and register it in the "Top" site, using a special "add site" API.  Then, configure the "Bottom" site to use the "Top" Keystone.

The Full Picture

Here is how the entire system looks:

Is it really that simple?

From the user experience point of view - We hope it is.  
But in order to get it there, we needed to handle quite a few obstacles:

• Resource Synchronization across the Multi-site
• Cross-site network
• Image synchronization
• Metadata synchronization (e.g. flavors)
• Resource Status monitoring and propagation (i.e. so that you can see what's happening from the "Top" dashboard) 

Coming next

In our coming posts, we will dive into resource synchronization, the cross-site networking, explain how we tackled the status and notification updates and share our approach to meeting large scale deployments.

Please share your thoughts about this in the comments.
We will be talking about this project in the upcoming OpenStack Tokyo summit, so if you're coming there, be sure to attend our talk.

To join the development effort: 

• Homepage: https://launchpad.net/tricircle
• Source: https://github.com/stackforge/tricircle 
• Blueprints - https://blueprints.launchpad.net/tricircle
• IRC Channel #openstack-tricircle

Centralized vs. Distributed SDN Control Path Paradigm

In a previous post I covered some misconceptions about SDN management and my view of the importance of the hybrid(proactive/reactive) model for scalability, and how we used this approach in the design of Dragonflow.

Today, i will discuss the two existing paradigms to SDN control path, and how it affects our road map in the Dragonflow project.

The first approach is the Centralized Controller.  Here, a controller (or a cluster of them) manages all the forwarding elements in the system, and retains a global view of the entire network. 

Most SDN controllers today run this way (ODL, ONOS, as well as Kilo version of Dragonflow).

The second approach is the Distributed Control Path.  Here, a local controller runs on each compute node and manages the forwarding element directly (and locally).  Thus, the control plane becomes distributed across the network.  

However, the virtual network topology needs to be synchronized across all the local controllers, and this is accomplished by using a distributed database.

Like everything else in life, there are advantages and disadvantages to each approach.  So, let's compare: 

Centralized Control Path

Pros

• The controller has a global view of the network, and it can easily ensure that the network is in a consistent, optimal configuration
• Simpler, agentless solution - Nothing needs to be installed on the compute nodes 
• Any and all southbound APIs can be supported directly from the centralized controller (easier to integrate with legacy equipment)

Cons

• Added latency in newly established flows, becomes a bottleneck in large scale deployments
• Dependency on the controller cluster availability and scale
• All advanced services are handled centrally, instead of locally, perpetuating a bottleneck as the scale grows
• Large scale is usually controlled via BGP-based Confederations and multiple SDNC clusters which add more latency and complexity

Distributed Control Path

Pros

• You can manage policies and introduce advanced services locally on each compute node, since you already have a local footprint
• Significantly better scalability, now that you have the control plane completely distributed
• Significantly better latency during reactive handling of PACKET_IN
• Highly-available by design and no single-point-of-failure
• Easier to integrate Smart NIC capabilities on local host level

Cons

• Synchronization of the virtual network topology can be a challenge, as the amount of compute nodes increases
• No global view
• Extra compute is done on the local host
• If you have heterogeneous forwarding elements (e.g. legacy switches), you need to have a centralized controller that connects them to the distributed control plane (which can complicate the management) 

What we chose and why

We decided earlier-on to go with a hybrid reactive/proactive model (against the widely-accepted proactive-only approach), as we saw that its advantages were overwhelming.  

The winning point of the reactive mode, as we see it, is that it improves the performance of the datapath, taking the performance toll instead on the control path of newly-established flows.  The main reason for that being a dramatic reduction of the number of flows that are installed into the forwarding element.

When combined with a pipeline that is deployed proactively, we could maximize the benefits of the reactive approach, while minimizing its cost.

However, like all solutions, at a certain scale it will break.

In very large deployments (e.g. a full datacenter), a central controller cluster becomes overwhelmed with the increase in volume of new connections.

A centralized controller made sense, while we were only handling new L3 connection path establishment (Dragonflow in Kilo).

However, when we came to add reactive L2 and other advanced services (like DHCP, LB, etc.), we realized that scaling the centralized controller cluster was becoming a huge challenge. 

A different approach must now be taken, and we believe this approach is to place a local controller on each compute node.  

Now that the control path bottleneck was mitigated, the problem moves to the logical data distribution between all the local controllers.

To mitigate that, we believe we can reuse the reactive approach, by letting the local controllers synchronize only the data they actually need (in lazy mode) use key/value distributed db engines that provide low latency. 

Sure, this will probably take some performance toll on the establishment of new flows, but we believe it will dramatically reduce the amount of data synchronization required, and therefore will take us to the next scalability level.

What we are doing in Dragonflow

We are currently working on introducing reactive L2 and L3 model into the local Draongflow controller.  

We are creating a pluggable distributed database mechanism to serve the logical data across the datacenter, which will enable the user to chose the best database to meet their specific needs and scale.

As always, we would love to have more people join the Dragonflow community and contribute to its success.

For additional information about the pluggable database layer, you can check out Gal Sagie's new blog post.

DragonFlow SDN based Distributed Virtual Router for OpenStack Neutron

In my previous posts I presented a PoC implementation of an embedded L3 controller in Neutron that solves the network node bottleneck the SDN way, and overcomes some of the DVR limitations.

In this post I am going to cover the first release of DragonFlow - an SDN based Distributed Virtual Router L3 Service Plugin (aligned with OpenStack Kilo), now an official sub-project of Neutron.

The main purpose of DragonFlow is to simplify the management of the virtual router while improving performance, scale and eliminating single point of failure together with the network node bottleneck.

The DrangonFlow solution is based on separation of the routing control plane from the data plane. This is accomplished by implementing the routing logic in distributed forwarding rules on the virtual switches (called "flows" in OpenFlow terminology). To put this simply, the virtual router is implemented using OpenFlow flows only.

DragonFlow eliminates the use of software stack to act as a virtual router (the Linux network namespaces in the DVR and the legacy L3 architecture), it use purely OVS flow to act as a virtual router. A diagram showing DragonFlow components and overall architecture can be seen here:

DragonFlow High-level Architecture

DragonFlow Features for the Kilo Release:

• East-West traffic is fully distributed using direct flows, reactively installed upon VM-to-VM first connection.

• Support for all ML2 type drivers GRE/VXLAN/VLAN

• Support for centralized shared public network (SNAT) based on the legacy L3 implementation

• Support for centralized floating IP (DNAT) based on the legacy L3 implementation

• Support for HA, in case the connection to the Controller is lost, fall back to the legacy L3 implementation until recovery. Reused all the legacy L3 HA. (Controller HA will be supported in the next release).

Key Advantages:

• Performance improvements for inter-subnet network communication by removing the number of kernel layers (namespaces and their TCP stack overhead)
• Scalability improvements for inter-subnet network communication by offloading L3 East-West routing from the Network Node to all Compute Nodes
• Reliability improvements for inter-subnet network communication
• Simplified virtual routing management, mange only active flows and not all possible 
• Non Intrusive solution, does not rely on ML2 modification

How it works

1. On bootstrap the L3 service plugin sends an RPC message to the L2 service plugin, setting the L3  Controller Agent as the controller of the integration bridge.
2. The Controller queries the OVS  for its port configuration via Openflow, matches the actual ports configured on the OVS to the Neutron tenant networks data model. 
3. Then, it installs the bootstrap flow pipeline that offloads all L2 traffic and local subnet L3 traffic into the NORMAL pipeline, while sending every unmatched VM-to-VM inter-sunbet traffic to the controller.

DragonFlow L3 Service Bootstrap

The following diagram shows the multi-table OpenFlow pipeline installed onto the OVS integration bridge (br-int) in order to represent the virtual router using flows:

bootstrap flows pipeline

The base table pipe-line is installed proactively on bootstrap while the East-West rules on the L3 Forwarding table are installed reactively on each first VM-to-VM communication.

If you would like to try it yourself, install guide is available here.

To join the development effort: 

• Homepage: http://launchpad.net/dragonflow
• Source: http://git.openstack.org/cgit/stackforge/dragonflow
• Bugs: http://bugs.launchpad.net/dragonflow
• Blueprints - https://blueprints.launchpad.net/dragonflow

My next post will cover the L3 reactive OpenFlow application, and how we install the East-West reactive flows. 

Neutron DVR The SDN Way

In my previous posts I covered the existing Openstack virtual L3 implementations, from the base centralized implementation to the state of the art solution, DVR, that distributes the load among the computes nodes.

I also summarized the limitations of DVR and hopefully convinced you of the motivation for yet another L3 implementation in Neutron.

Just to quickly recap, the approach for DVR’s design was to take the existing centralized router implementation based on Linux network namespaces and clone it on all compute nodes. This is an evolutionary approach and a reasonable step to take in distributing L3 from which we've learned a lot. However, this solution is far form optimal.

In this post I’m going to present a PoC implementation of an embedded L3 controller in Neutron that solves the same problem (the network node bottleneck) the SDN way, and overcomes some of the DVR limitations.

I chose to start with just the East-West traffic, and leave the North-South for the next step.

I will share some of my ideas for North-South implementation in upcoming posts, as well as links to the Stackforge project.

If you're interested in joining the effort or take the code for a test drive, feel free to email me or leave a comment below.

SDN Controllers are generally perceived as huge pieces of complex code with hundreds of thousands of lines of code that try to handle everything.  It's natural that people would not deem such software to be capable of running "in line".  However, this is not necessarily so.  As I will demonstrate, the basic SDN Controller code can be separated and deployed in a lean and lightweight manner.

It occurred to me that by combining the well-defined abstraction layers already in Neutron (namely the split into br-tun, br-int and br-ext) and a lightweight SDN controller that can be embedded directly into Neutron, it will be possible to solve the network node bottleneck and the L3 high-availability problems in the virtual overlay network in a simple way.

Solution overview 

The proposed method is based on the separation of the routing control plane from the data plane. This is accomplished by implementing the routing logic in distributed forwarding rules on the virtual switches. In OpenFlow these rules are called flows. To put this simply, the virtual router is implemented in using OpenFlow flows. 
The functionality of the virtual routes that we should address in our solution is:

• L3 routing
• ARP response for the router ports
• OAM, like ICMP (ping) and others

L3 Controller embedded in Neutron POC

We decided to create our controller PoC with Openstack design tenets in mind. Specifically the following:

1. Scalability     - Support thousands of compute nodes
2. Elasticity       - Keep controllers stateless to allow for dynamic growth 
3. Performance  - Improve upon DVR
4. Reliability      - Highly available
5. Non intrusive - Rely on the existing abstraction and plug-able module

What we intend to prove is:

• We indeed simplify the DVR flows
• We reduce resource overhead (e.g. bridges, ports, namespaces)
• We remove existing bottlenecks (compared to L3Agent and DVR)
• We improve performance

We've started benchmarking the different models and I'll post results as soon as we have them ready.

Reactive vs. Proactive Mode

SDN defines two modes for managing a switch: reactive and proactive.
In our work, We decided to combine these two modes so that we could benefit from the advantages of both modes (although if you believe the FUD about the reactive mode performance, it is quite possible to enhance it to be purely proactive). To learn more about this mixed mode see my previous blog

The Proactive part

In the solution we install a flow pipeline in the OVS br-int in order to offload the handling of L2 and intra-subnet L3 traffic by forwarding these to the NORMAL path (utilizing the hybrid OpenFlow switch). This means that we reuse the built in mechanisms in Neutron for all L2 traffic (i.e. ML2 remains untouched and fully functional) and for L3 traffic that does not need routing (between IPs in the same tenant + subnet).

In addition we use the OVS OpenFlow extensions in order to install an ARP responder for every virtual router port. This is done to offload ARP responses to the compute nodes instead of replying from the controller.

The Reactive part

Out of the remaining traffic (i.e. inter-subnet L3 traffic) the only traffic that is handled in a reactive mode is the first packet of inter-subnet communications between VMs or traffic addressed directly to the routers' ports.

The Pipeline

Perhaps the most important part of the solution is the OpenFlow pipeline which we install into the integration bridge upon bootstrap.
This is the flow that controls all traffic in the OVS integration bridge (br-int).
The pipeline works in the following manner:

1. Classify the traffic
2. Forward to the appropriate element:
1. If it is ARP, forward to the ARP Responder table
2. If routing is required (L3), forward to the L3 Forwarding table (which implements a virtual router)
3. Otherwise, offload to NORMAL path

At the end of this post is a detailed explanation of the pipeline and implementation brief in case you're interested in the gory details.

In my following posts I will cover in detail the code modifications that were required to support the L3 controller as well as publish a performance study comparing the L3 Agent, DVR and the L3 controller for inter subnet traffic.

If you want to try it out, the code plus installation guide  are available here.

Again, if you'd like to join the effort, feel free to get in touch.

detailed explanation of the pipeline

The following diagram shows the multi-table OpenFlow pipeline installed into the OVS integration bridge (br-int) in order to represent the virtual router using flows only:

L3 Flows Pipeline

The following table describes the purpose of each of the pipeline tables:

ID	Table Name	Purpose
0	Metadata & Dispatch	Tag traffic with the appropriate metadata by input port, the value is the segmentation ID Input traffic from the tunnel bridge is offloaded to the NORMAL path All other traffic is sent to the "Classifier" table for classification
40	Classifier	All ARP traffic is sent to the ARP responder table All broadcast and multicast traffic is offloaded to the Normal path All L3 traffic is sent to the L3 forwarding table
51	ARP Responder	This table handles ARP responses for the virtual router interfaces. The vlan tag is removed from all other traffic and offloaded to the NORMAL path
52	L3 Forwarding	This table is used by the Controller to install flows to handle VM to VM inter subnet traffic (detailed explanation below) Traffic destined for a virtual router is sent to the controller (e.g. ping, packet_in, etc) All local subnet traffic is offloaded to the NORMAL path All other traffic is sent to the public network table
53	Public Network	will be described in following blogs

PoC Implementation brief

For the first release, we used the Open vSwitch (OVS), with OpenFlow v1.3 as the southbound protocol, using the RYU project implementation of the protocol stack as a base library.
The PoC support Route API for IPV4 East -West traffic.  In the current PoC 
implementation, the L3 controller is embedded into  the  L3 service plugin.  

Common Misconceptions about SDN Controller Management and Scalabilty

Here are a few common misconceptions about SDN controller management and scalability.

#1: Either Proactive mode or Reactive mode

When an SDN controller wants to enforce a set of rules or a policy on a forwarding element, it uses the southbound API, OVSDB, OpFlex, OpenFlow or others. In OpenFlow this process is called flow installation and it can be done using different methods: proactive, reactive or mixed (aka hybrid approach).  

Proactive

During bootstrap, the controller installs all flows and pipelines (multi-table entries) into all the forwarding elements.  The flows must cover all possible scenarios.

Whenever there is a change in the network, the controller removes or installs flows where necessary.

Reactive

When a packet arrives in a switch, a look-up is performed on the flow tables.
If there is no match and the switch is connected to a controller, it will forward the packet (either with or without its payload) to the controller.
In OpenFlow, this is called a PACKET_IN message.

It is also possible to create rules that forward the packet to the controller when matched and we call that...

Mixed Proactive / Reactive

With the mixed approach, you can benefit from the best of both worlds, achieving a balance between dynamic management and performance.

For frequently-used and rarely-modified flows, you install a pipeline and flows proactively.

For unmatched flows, or for flows that you want to handle in-line in your SDN application, you install a flow to forward the packet to the controller.

This mixed approach allows the controller to focus on making real-time dynamic decisions only on the traffic that requires it (reactive), while leaving the heavy-lifting of the majority of the traffic to the real-time forwarding element (proactive).

In addition, with this approach you can avoid over-provisioning the forwarding element with flows that are rare, thus dramatically reducing the number of entries in the flow tables. 

#2: Lack of Scalability and High Availability

In SDN, the control plane (i.e. the SDN Controller) is separated from the data plane (i.e. the Forwarding Elements).

Due to the centralized nature of control in SDN, we now need to support high availability and redundancy of the Control Plane.

In OpenFlow, the Forwarding Element (the Switch) connects to the controller via TCP or via TLS for secure channels.

Up until OpenFlow v1.2, whenever the connection to the controller was lost, the switch would lose the ability to forward PACKET_IN messages to the controller, thus having to drop all unmatched traffic or handle it in the NORMAL switch (only in switches that implemented the dual nature). 

This non-deterministic approach would yield unpredictable network behavior 
while the controller was unavailable.

OpenFlow v1.2 introduced the capability of working with multiple controllers.
The Master and Slave modes provide a mechanism for Active-Passive high availability, whereas the Equal mode provides an Active-Active model.

OpenFlow multiple controllers

With the multi-controller capability, we gained the control plane high availability we needed, improved reliability, fast recovery from failure and controller load balancing.

#3: You cannot design SDN Applications for Really Big Scale

Utilizing the two mechanisms I covered - Mixed-Mode and Multi-Controller - is the key to designing really big scale SDN applications.

If you take care to design your application to be stateless (whenever possible) and share nothing (or little) with other controller instances you can benefit from the Reactive mode where any instance can handle any flow.

In my opinion, the best approach for achieving minimal controller response latency and maximal bandwidth while keeping the dynamic allocation of flows, is by using all of these mechanisms together.

When you add to that the OVS OpenFlow extensions and the dual-nature Hybrid OpenFlow capability (which I covered in my previous post), you can really gain a dramatic performance and management boost.


In my next post I will demonstrate how we utilized these design guidelines and capabilities in a prototype for an alternative to Neutron's L3 Distributed Virtual Router.

Hybrid OpenFlow Switch

In my last post I summarized the DVR solution and tried to explain the motivation for yet another L3 implementation in Neutron that I am going to present in the coming posts.

This 2-post series is intended to cover basic SDN and OpenFlow mechanisms that we used in the L3 controller:

• Hybrid OpenFlow Switch 
• SDN models for managing the forwarding elements (switches)

Hybrid OpenFlow Switch

Hybrid OpenFlow switch was introduced in OpenFlow/1.1. Hybrid switches support both OpenFlow operation pipeline and normal (legacy) Ethernet switching functionality.

The hybrid switch allows forwarding of packets from the OpenFlow

pipeline to the normal pipeline through the NORMAL and FLOOD reserved ports.

The main reason for introducing the hybrid switch was to optimize the handling of operations like MAC learning, where a reactive approach was just not efficient - Doing MAC learning in the OpenFlow controller poses a significant cost in terms of network bandwidth and latency and does not scale for large networks.

The NORMAL action comes to the rescue and lets us offload legacy non-OpenFlow pipeline (like MAC learning mechanism, VLAN, ACL, QoS and other base features) to the forwarding element kernel module, which is optimized to handle such operations in near-line-rate.

But what happens when the NORMAL action is used in an OpenFlow flow?

Basically, what happens is that the traffic is redirected to a completely separated processing pipeline. This is illustrated in the diagram below.

OVS Hybrid OpenFlow Switch Pipelines

The OpenFlow pipeline and the Normal Pipeline, each act as a completely isolated switch.

There are, however, some issues with this hybrid approach.

• NORMAL pipeline is not standardized, so it behaves differently on switches from different vendors - There is a variance in the supported features, and no standard way to configure them
• NORMAL pipeline does not play well with some OpenFlow actions, for example if the port is tagged for the NORMAL pipeline (using the ovs-vsctrl), you can not tag it using OpenFlow actions and then forward it to the NORMAL path, because it will end-up dropped due to double tagging error.

The Open virtual Switch extensions to OpenFlow were developed to support these extra features using Flows, for example the LEARN action (Open vSwitch extension to OpenFlow) for MAC learning .

In my next post I will cover the SDN models for managing the forwarding elements in reactive, proactive and mix modes.