Adam Raffe Cloud, Data Centre & Networking

Say Hello to Azure Container Instances (ACI!)

It’s been some time since I last posted a blog - I’ve been spending the majority of my time settling in to my new role at Microsoft and learning as much as I can about the Azure platform. Considering my background, it seems fitting that my first “Azure related” blog post is all about….ACI! In this case however, ACI stands for Azure Container Instances - a new Azure offering announced this week and currently in preview. So what are Azure Container Instances?

Let’s first consider how we might run containers in the cloud today. Up until this point, if you wanted to run containers on Azure, you had two main options:

a) Provision your own virtual machines and run containers on top, or

b) Use Azure Container Service (ACS) - this service essentially allows you to quickly provision a cluster using either Kubernetes, DC/OS or Docker Swarm as the orchestration engine.

Both of these solutions suffer (to varying degrees) from the overheard of having to manage the underlying VM infrastructure and everything that goes along with that - maintenance, security, patching and so on. Wouldn’t it be nice if we could simply create our containers in the cloud without having to worry about the infrastructure they are running on? Look no further than Azure Container Instances.

ACI instances give us the ability to run containers easily from the command line or using ARM templates, with no need to create the underlying VMs. In addition, these container instances are billed _per second, _meaning they should work out to be extremely cost-effective. The official MS page for ACI is here.

To see how ACI works, let’s try it out. To start with, I’m going to create a single container using the Azure CLI. I’ll use the following command:

az container create -g ACI --name acitest --image nginx --ip-address public

This command is essentially asking for a container to be created under the resource group named “ACI”, with a name of “acitest”, using the public nginx image (which will be pulled from Docker Hub). The command also asks for a public IP address to be assigned to the container. The results of this can be seen in the following figure:


The JSON output that results shows a wealth of information about the container - for example, I can see the amount of CPU and memory (1 CPU core, 1.5Gb of RAM - this is configurable), the IP address that has been assigned as well as some information about the resource group. You can also see that the provisioning state is “creating”. If I run the command _az container list _a few seconds later, I can see the container state is now “succeeded”:


At this point, I can browse to the public IP address that Azure assigned to my container and I receive the default nginx page.

I can also get the log output from the container using the command az container logs:


OK, so this is all very nice - but wouldn’t it be great if we could automate this process? And it would be even better if we could create multiple container instances at the same time. It is of course possible to do this using ARM templates. Check out the example template below (full template available here):


The JSON shown above should do the following:

  • Create a new container group named “aciMultiGroup”.

  • Create two container instances - “aci-container1” and “aci-container2” as part of the group.

  • Expose port 80 from container 1 and assign a public IP address to the group.

This template introduces us to the container group concept. The idea behind this is that one or more containers with similar requirements or lifecycles can be deployed and managed together inside a container group (somewhat similar to the ‘pod’ concept in Kubernetes).

So let’s deploy this template and see what happens:


OK, everything looks good. The output of the script has given me the public IP address of the container group - let’s try and browse to that:


Success! Hopefully this post has given you a taster of what Azure Container Instances can do and the potential they have. Thanks for reading!

Moving on

A little over three years ago, I was introduced to Cisco’s Application Centric Infrastructure (ACI) for the first time. For someone who had spent many years at Cisco working on “traditional” networking platforms, this was something of a revelation - the ability to define network connectivity in a programmable manner using a policy based model was a major departure from anything I had done before. Since then, I’ve been lucky enough to work with a variety of customers around the globe, helping them to design and deploy the ACI solution. I’ve been part of a great team of people, worked closely with the INSBU team (responsible for ACI) and presented to hundreds of people at Cisco Live.

Over the last few months, I’ve spent some time thinking about what I do next: as ACI becomes more mainstream, do I continue with more of the same - expanding my skill set to include the other great products (Tetration, Cloudcenter,etc) that Cisco has in the data centre - or do I take a slightly different path? After some serious consideration, I’ve decided to go with the latter option - later this month, I’ll be joining Microsoft as a Cloud Solutions Architect, working with the Azure platform.

I’ve thoroughly enjoyed writing this blog over the last few years and want to thank everyone who has read the posts, commented or given me feedback. I’m hoping to continue blogging occasionally, so keep an eye out for the odd Azure-related post!


Learning ACI - Part 12: Inter-VRF and Inter-Tenant Communication

ACI has the ability to divide the fabric up into multiple tenants, or multiple VRFs within a tenant. If communication is required between tenants or between VRFs, one common approach is to route traffic via an external device (e.g. a firewall or router). However, ACI is also able to provide inter-tenant or inter-VRF connectivity directly, without traffic ever needing to leave the fabric. For inter-VRF or inter-tenant connectivity to happen, two fundamental requirements must be satisfied:

  1. Routes must be leaked between the two VRFs or tenants that need to communicate.

  2. Security rules must be in place to allow communication between the EPGs in question (as is always the case with ACI).

The question is, what is the correct way to configure this type of connectivity? I’ve seen quite a bit of confusion around this, particularly when it comes to deciding where to configure subnets (at the bridge domain level or EPG level), so hopefully I can provide a bit of clarity in this post. I’m going to cover three main scenarios:

  1. Inter-VRF communication, where there is a 1:1 mapping between bridge domains and EPGs within the VRFs.

  2. Inter-tenant communication, where there is a 1:1 mapping between bridge domains and EPGs.

  3. Inter-VRF communication, where multiple EPGs are associated with a single bridge domain in one or both of the VRFs.

Scenario 1: Inter-VRF Communication - 1:1 Mapping Between BD and EPG

In this example, we have a simple setup with two VRFs - each VRF has a single bridge domain (BD) and End Point Group (EPG) configured as shown in the following diagram:


At the beginning of this article, I mentioned that there were two requirements when configuring inter-VRF / inter-tenant communication: leaking of routes and security rules. ACI uses contracts to control both of these.

Let’s deal first with how we leak routes from the consumer side (EPG2 in our example) to the provider side (EPG1). To make this happen, we do the following:

  1. Configure our subnet at the BD level and ensure it is marked with Shared Between VRFs.

  2. Create a contract that defines the appropriate port protocols we wish to allow and make sure the scope is defined as Tenant.

  3. Attach the contract to the provider and consumer EPGs.

At this point, we have something like this configured:


As you can see, associating the contract with the provider and consumer EPGs within each VRF results in the consumer side BD subnet ( being leaked to VRF A.

OK, so that’s great - we have routes leaked between VRFs in one direction (from consumer to provider). But clearly nothing is going to work until we leak routes in the opposite direction. So how do we get those routes leaked? This is the bit where a few people get tripped up.

The answer is that - on the provider side - we need to configure our subnets at the EPG level, rather than the bridge domain level (as is the case in ‘normal’ deployments that don’t involve route leaking). This is shown in the following diagram:


There is a long technical explanation as to why this is necessary - I won’t bore you with that here, but fundamentally, the fabric has to be told which subnet belongs to which provider EPG, which is why the subnet must be configured at the EPG level rather than the BD level.

One point to note here is that you do not need to export contracts when configuring inter-VRF communication. This is only a requirement for inter-tenant communication.

Hopefully that’s clear so far - let’s move to our second scenario.

Scenario 2: Inter-Tenant Communication - 1:1 Mapping Between BD and EPG

In this scenario, we are going to configure communication between EPGs which sit in different tenants, as shown here:


In terms of configuration, this scenario is actually very similar to scenario 1 - we still need to configure the following:

  • Consumer subnet at the bridge domain level (marked as S_hared Between VRFs_).

  • Provider subnet at the EPG level (marked as S_hared Between VRFs_).

The major difference in this scenario is that we must now configure the scope of the contract as Global, and we must also export the contract from the provider side tenant to the consumer side tenant. On the provider side, export the contract under the Security Policies section of the tenant config. On the consumer side, we will consume the exported contract as a Contract Interface under the EPG. The final configuration looks something like this:


OK, that was easy enough - onto our last scenario where things get a little more complex.

Scenario 3: Inter-VRF Communication - Multiple EPGs Associated With a Single BD

For this last scenario, we are again looking at inter-VRF communication, but this time we have more than one EPG associated with a single bridge domain in both VRFs, as shown here:


If we follow our previous examples, we would configure our consumer side subnets at the BD level and our provider side subnets at the EPG level. Hold on a minute though, there’s an issue with that - if we assume that we have only a single subnet on the provider side and that we are using that subnet for all EPGs, does that mean we have to configure the same subnet on all provider EPGs? Well actually, we can’t do that - subnets configured on an EPG must not overlap with other subnets in that same VRF. So how do we get around this?

The answer to this conundrum is that we need to return to a configuration where the subnet is configured at the BD level on both provider and consumer sides. However, one of the consequences of this is that we then need to create a bidirectional contract relationship between the EPGs - in other words, we must provide and consume the contract from both sides in order to allow route leaking to happen. This ends up looking like this:


Now, if you are familiar with ACI and the way in which rules are applied in hardware, you may spot a downside to this approach. The problem with consuming and providing a contract in both directions is that this means double the number of security rules are programmed in hardware (TCAM). If you are a heavy user of contracts in your environments, doubling up on the number of rules may be a concern for you.

If you are concerned about this doubling of rules, there is a potential solution. In order to reduce the number of rules that we need to program, we can configure the contracts that we use to be unidirectional in nature. What I mean by this is that we can un-check the boxes entitled Apply Both Directions _and _Reverse Filter Ports when we create the contract. Now, having those boxes checked is generally important for allowing return traffic back to the initiating host (on the consumer side). So if we un-check them, how do we allow return traffic?

The answer is that we configure two contracts - one that allows traffic to be initiated (let’s say it allows any port –> port 80) and one allowing the reverse traffic (let’s say port 80 –> any port). Those contracts are then applied separately to the consumer and provider side EPGs, as shown in the following two diagrams:



By doing this, we cut the number of entries required in the hardware required by half compared to the solution shown in the first part of this scenario.

Hopefully this helps you out a bit if you are configuring this feature - thanks for reading.

Learning ACI - Part 11: Transit Routing

The 1.1(1j) & 11.1(1j) release of ACI introduced support for transit routing. Prior to this, the ACI fabric acted as a ‘stub’ routing domain; that is, it was not previously possible to advertise routing information from one routing domain to another through the fabric. I covered L3 Outsides in part 9 of this series where I discussed how to configure a connection to a single routed domain. In this post, we’ll look at a scenario where the fabric is configured with two L3 Outsides and how to advertise routes from one to another. Here is the setup I’m using:


In my lab, I have a single 4900M switch which I have configured with two VRFs (Red and Green) to simulate the two routing domains. In the Red VRF, I have one loopback interface - Lo60 ( which is being advertised into OSPF. In the Green VRF, I have Lo70 ( and Lo90 ( which are also being advertised into a separate OSPF process.

On the ACI fabric side, I have two L3 Outsides which correspond to the two VRFs. These two L3 Outsides are associated with a single private network (VRF) on the ACI fabric. OSPF is configured on both L3 Outsides with regular areas.

At this point, my OSPF adjacencies are formed and my ACI fabric is receiving routing information from both VRFs on the 4900M, as can be seen in the following output (taken from the ‘Fabric’, ‘Inventory’ tab and then under the specific leaf node, under the Protocols and then OSPF section):

Screen Shot 2015-10-17 at 17.38.33

You can see from this output that the fabric is receiving all of the routes for the loopback addresses configured under the Red and Green VRFs on the 4900M. Let’s now take a look at the routing table for the Red VRF on the 4900M:

Screen Shot 2015-10-17 at 17.42.40

It’s clear from the above output that the Red VRF is not receiving information about either the or prefixes from the Green VRF - in other words, the ACI fabric is not currently re-advertising routes that it has received from one L3 Outside to the other.

Let’s say I now want to advertise the prefixes from the Green VRF ( and into the Red VRF - how do I enable that? The main point of configuration for transit routing is found under ‘External Networks’ under the L3 / Routed Outside configuration. The key here is the Subnets configuration - in previous versions of ACI, this has been used only to define the external EPG for policy control between inside and outside networks. Now however, the subnets configuration is used to also control transit routing in addition to policy control. Here are the options and what they are used for:

  • Security Import Subnet: Any subnet defined as a security import subnet is used for policy control. Effectively, a subnet defined in this way forms the external EPG - this is the same functionality that existed in previous ACI releases. If I define a subnet as a security import subnet, this subnet will be accessible from internal EPGs (or other external EPGs), as long as a suitable contract is in place. Importantly, this option has nothing whatsoever to do with the control of routing information into or out of the fabric.

  • Export Route Control Subnet: This option is used to control which specific transit routes are advertised out of the fabric. If I mark a subnet with export route control capability, I am telling the ACI fabric that I want those routes to be advertised to the external device. Note that this option controls the export of transit routes only - it does not control the export of internal routes configured on a bridge domain (you can see evidence of this in the 4900M routing table output above, in which you can already see one of my BD routes - in the table).

  • Import Route Control Subnet: Similar to the export route control option, this option can be used to control which routes are allowed to be advertised into the fabric. Note that import route control is currently only available if BGP is used as the routing protocol.

How are these used in practice? Let’s start with a simple example. I’m going to advertise just one of my ‘Green’ subnets (let’s say towards the ‘Red’ VRF. To do this, I add the subnet to the L3 Out facing the Red VRF and mark it with the ‘export route control’ option:

Screen Shot 2015-10-17 at 19.33.01

Now if I check the ‘Red’ routing table, I see the prefix advertised from the fabric:

Screen Shot 2015-10-17 at 19.34.37

If I add the prefix to my subnets list, that transit route will also be advertised from the fabric to the Red VRF.

You might now be wondering how you would handle a large list of transit routes; would they all need to be individually entered into the subnets list? No - you can use the Aggregate Export option. This option is currently only available when “” is used as the subnet; essentially, this option tells the fabric to advertise all transit routes. Checking the ‘aggregate’ option is important here - if you simply enter “” as an export route control subnet without the aggregate option, the fabric will try and advertise the route only. In my example, I’ve now removed the individual subnet and entered with the aggregate option:

Screen Shot 2015-10-17 at 19.39.56

I now see both my subnets advertised to the Red VRF:

Screen Shot 2015-10-17 at 19.41.12

So that’s the routing taken care of - but there’s an additional step if you want traffic to flow. Remember, the ACI fabric models external destinations as EPGs. If you have two external destinations that need to communicate through the fabric, you must have both of those external destinations covered by a Security Import Subnet. As an example, if I wanted to allow hosts on (part of the Red VRF) to talk to hosts on (Green VRF), in addition to exporting the routes themselves (in both directions), I would need to define both of those subnets with the security import option:

4900M-External Subnets Configuration:

Screen Shot 2015-10-17 at 19.45.38

4900M-External-2 Subnets Configuration:

Screen Shot 2015-10-17 at 19.47.57

I would then need to provide / consume contracts between these networks for traffic to flow. In the above example, I could have just created a single subnet on each side and marked it with export route control, aggregate export and security import options - that would effectively allow all routes and all destinations to communicate with each other (assuming contracts).

Thanks for reading!

Learning ACI - Part 10: Programming the ACI Fabric

Everything I’ve shown so far in this blog series has been focused on using the APIC GUI to achieve what we need. This is fine for small environments or for becoming familiar with ACI, but what if you need to configure 100 tenants, each with 50 EPGs, tens of private networks and bridge domains, multiple contracts to allow traffic and a L3 Outside or two thrown in? Configuring all of that through the GUI is clearly going to be massively time consuming and probably quite error prone, so we need to find a better way of doing things. ACI has a wide variety of programmability options that can be used to automate the provisioning of the fabric and its policies.

One particularly strong aspect of ACI is how easy it is for network engineers to start interacting with the fabric in a programmatic manner. Speaking from personal experience, it’s incredibly easy for someone with a ‘traditional’ network engineering background with limited development experience to start using the programmability features of ACI. With that said, let’s take a look at some of the options we have around ACI programmability:

  • Native REST API using XML or JSON

  • ACI Python SDK (“Cobra”)

  • ACI Toolkit

Let’s take a look at the native REST API option first.


The first option - and arguably the one most people look at first - is to interact with the APIC API using raw XML or JSON. Using raw XML / JSON has the advantage of not requiring the user to have any real programming knowledge (for example, Python) and interacting with the REST API natively can be done through a variety of tools (e.g. any tool capable of posting data to a URL such as cURL, Postman plugin for Chrome, etc). So how do we get started with this?

Before you start, it’s worth bearing in mind that to use the REST API effectively, it’s useful to have some understanding of how the object model works. For example, if we want to create a tenant, private network, bridge domain and EPG using the API, we need to have a bit of knowledge about how these objects relate to each other. Fortunately, the APIC provides some tools to help us understand this - one of which is the object store browser (Visore).

To get to Visore, browse to /visore.html and log in. Once in, you'll be able to query for classes or objects using the search box at the top of the screen. In this example, I'm going to search for all instances of class _fvTenant:_


In the above screenshot, my query has returned 32 objects of class fvTenant (I have 32 tenants on my fabric), two of which I have shown here. From here, you can see the distinguished name of each tenant object (e.g. uni/tn-LAN01). Now, the really useful part of this is that clicking the right arrow (>) next to the distinguished name will take you to another screen showing all of the child objects of the tenant object in question:


In the above screenshot, you can see two of the child objects for my tenant (named “adraffe-test”) - specifically, we see an application profile (class fvAp, named “Test-App-1”) and a bridge domain (class fvBD, named “Adam-Test-BD”). Scrolling further down reveals additional objects such as private networks, contracts and filters. Clicking the right arrow next to the DN takes you further down the tree - for example, drilling further into the “fvAp” object here would take me to the EPG objects contained within.

Now that we have some knowledge of the object model, we can start with some basic interaction with the REST API. The first thing I need to do is authenticate, which I can do by posting to the URL in the screen shot below and sending the XML shown:


Now that I’ve authenticated, I can start creating some objects through the ACI. I’ll start by creating a new tenant named ‘Finance’:


That’s it - sending that simple piece of XML to the APIC created a new tenant. Now let’s create a private network, bridge domain, application profile and EPG inside my new tenant. I post to the same URL as in the last example (the “uni” in the URL is used to represent “policy universe”), but this time I send the following:

Tenant XML

Let’s take a closer look at what the above XML does. Firstly, I need to include the parent tenant object I want to modify (Finance). I then create a new private network (fvCtx) named Finance-Net. Next, I create a bridge domain (fvBD) named Finance-BD, with a subnet address (fvSubnet) of I also create a relationship between my new bridge domain and the private network I created earlier (using fvRsCtx). Next up, I create an application profile (fvAp) called Finance-AP1 and add a new EPG (fvAEPg) to it named Finance-EPG. Finally, I create a relationship from my new EPG to the BD I created earlier (using fvRsBd).

This is extremely easy to get the hang of - you can use the API inspector and “Save As” tools which I referred to in my earlier post (as well as Visore) to help you to get familiar with the object model and build your own scripts.

ACI Python SDK (Cobra)

If you are familiar with Python, Cisco have a Python SDK (also known as “Cobra”) available. This SDK allows you to build a Python program to make API calls to the APIC without having to post raw XML or JSON as in the last example. You still need a working knowledge of the object model, which again the tools mentioned above can help with.

As a comparison, let’s take a look at the equivalent Python code which would create the same tenant, private network, bridge domain and EPG we created earlier (note that there is more to this script not shown here, such as importing the relevant modules, logging into the APIC, etc):


Anyone not familiar with Python may at this point be wondering if they should steer clear of the Python SDK - however, fear not: there is an extremely cool tool which will take the XML or JSON which you provide and automatically generate Python code for you. This tool is called arya (APIC REST Python Adapter) and is available here:

You can see this tool in action by simply downloading some XML / JSON from the APIC using the “Save As” feature and then feeding the file into the arya tool. The tool will generate the equivalent Python code for the XML or JSON you provided. Pretty cool, right?

ACI Toolkit

The final option I’ll look at here is the ACI Toolkit. One of the considerations for using either the native REST API or the Cobra SDK is that you do need to have some familiarity with the ACI object model. That’s not necessarily a huge mountain to climb, but some people may be looking for a quicker and simpler way of accessing the programmability features available with ACI. Step forward, ACI Toolkit.

The ACI Toolkit is essentially a set of Python libraries which takes the ACI object model and abstracts it into a greatly simplified version. Right now, the ACI Toolkit doesn’t provide full feature coverage in the same way that the native API or Cobra SDK does, but if you are looking for a simple way to create the most commonly used objects (EPGs, BDs, private networks and so on) then it’s worth taking a look at what the Toolkit can offer.

Here is an example of how the ACI Toolkit can be used to create some common objects within ACI:


One other nice bonus of the ACI Toolkit is that a number of applications are thrown in, such as the Snapback app (used for config rollback within ACI), End Point Tracker (used to keep track of and display end point information within the fabric), various visualisation tools and more.

You can get more info on the ACI Toolkit, including instructions to install here.

To sum up, hopefully it’s clear from this post that programmability in ACI is not limited to developers - anyone with a background in traditional networking should be able to get up to speed reasonably quickly and be able to start taking advantage of the benefits associated with programmability.

Thanks for reading.