Rick Mur

Limitless Networking

What is P4?

Recently I attended a workshop organised by the Open NFP organisation about Data Plane acceleration. The primary goal of the workshop was to get students and researchers (why was I there you may think) familiar with the P4 programming language.

P4 is a programming language created to simplify writing data planes for networking use cases.  Recently the P4-16 spec was released and could be considered a mature version of the language.

Now I’m not a hardcore developer. I know my way around in Python, GoLang and C#, but I never wrote anything more low level like C. P4 is created a little bit like GoLang, where I do not mean it as comparison, but as an architecture. P4 is designed so you only need to focus on the actual networking features that you want to make available on the hardware you are programming it for. Then when you compile it, it will generate runtime code for your hardware. Or as the creators explain it:

At one level, P4 is just a simple language for declaring how packets are to be processed. At another level, P4 turns network system design on its head.

There is no need to worry on low level memory management or other things that would slow down your development for that specific feature that nobody has in their platforms.

Now this is a major advantage. Full freedom in writing anything you’d like right? The downside is that very little (basic) functionality is not present when you start pushing code to a piece of hardware. You really have to do everything yourself. Which I think is  underestimated. P4 does help you in giving you some standard functionality in their spec that vendors should implement when they offer P4 capabilities, like basic switching or load balancing (ECMP) functionality.

Who supports it?

P4 is relatively new and therefore hardware support is not very common. I was able to find 2 products to support the full P4 spec at this moment.

Barefoot Tofino chip

Barefoot is the inventor of the P4 language and they have created a very fast ASIC like chipset. The chipset does not use a proprietary SDK like Broadcom, Cavium or silicon from large network vendors, but is fully open due to the support of P4.

They are partnering with white box vendors to create switches (32 port 100GBE) that are currently shipping.

Netronome Agilio SmartNIC

Netronome was one of the organizers of the workshop I attended and they have a shipping product. Their SmartNIC products are standard PCIe NIC’s with various connections that support offloading data plane functionality to the NIC rather than letting applications or the kernel spend a lot of CPU on it.

These products are not your typical switch or NIC, they basically do not have a default feature set available to use. The Barefoot switch for example, does not come with an operating system (NOS) and the NIC always needs to be programmed before any functionality works.

What about OpenFlow?

The most typically asked question would be: Is P4 the successor of OpenFlow? The answer to that is really simple: No! In some cases P4 may even use OpenFlow to program it’s forwarding table(s). Diagram courtesy of P4.org:

The P4 language is designed to program your networking hardware (or even software components). P4 does not contain a control-plane protocol to learn about forwarding table entries (that could be learned through a traditional or home grown routing protocol).  In other words: you have to write your own forwarding table programming protocol/application or use an existing one (like OpenFlow).

What about existing vendors?

There are rumours that major networking vendors (Cisco, Juniper, Nokia, etc.) could start supporting (parts of) P4 on their own silicon or implement the Barefoot Tofino chip in products, but no announcements have been made at time of this writing. Vendors are starting to open up their operating systems via programmability features (see Juniper’s Extension Toolkit JET for example), but this does not allow to directly program new features into the hardware (which is programmable in some cases like Juniper MX).


P4 will open up a lot of features that are currently unknown or don’t exist. It allows for a fully open and programmable data-plane. The use case for this technology would, as always, be the web-scale companies, who want to free up as much CPU resources to perform other tasks and leave the networking hardware to perform all networking functions.

For lower scale deployments, my opinion would be that, tools like DPDK, XDP or eBPF would already free up sufficient CPU resources and optimise user space or kernel networking enough for most people.

For more information on all the mentioned acronyms and tools, click the links in this blog.

One noticeable fun fact was that during the workshop I attended, I was able to spot one person on a Mac, two corporate Windows machines and all others running Linux 🙂

VMware NSX 101 – Components

Today I want to explain the basic components and the set-up of VMware NSX. In this case I’m referring to NSX for vSphere or NSX-V for short. I want to explain what components are involved, how you set them up for an initial deployment and what the requirements are.


At time of this writing the latest release is NSX 6.1.4. This version added support for vSphere 6, although you cannot use any vSphere 6 feature in this release, there is support for the platform itself only.


The first step is of course deploy your ESXi vSphere cluster with ESXi 5.5 or 6.0 with vCenter 5.5 or 6.0. I recommend using the vCenter Server Appliance (VCSA) instead of the Windows version. You will also need a Windows VM where the vSphere Update Manager is installed, this is not available as virtual appliance, only as Windows application. I also highly recommend installing an Active Directory server to manage all of your passwords. You will be installing a large amount of machines with all different usernames and possibly passwords. I recommend picking a very long and difficult one, as all VMware appliances seem to require a different password complexity.


After deploying the initial cluster it’s also essential to deploy a Distributed vSwitch across the nodes where you will be deploying virtual networks. For this feature you will need the Enterprise license (or the trial will suffice) when running ESXi 5.5 Update 3 and NSX (before that Enterprise Plus was required).

NSX Manager

The only component you download is the NSX Manager. This OVA file contains everything you will need for deploying NSX. After deploying the OVA through vCenter you are able to log-in to the Web GUI of the NSX Manager. The only setting available is connecting to the vSphere Single Sign-On server and to vCenter. When deploying all components make sure you set NTP servers for everything in the lab. There are many integrations and they all rely on authentication that also takes time into consideration.


After connecting the NSX Manager you will no longer need it that GUI. Within the vSphere Web Client a “Networking & Security” tab is now available.

NSX Controller

The next step is to deploy at least 3 NSX Controllers in the cluster. This is automatically done by the Manager after filling in required information. You need 3 for redundancy as the cluster will select a master controller, which is done by using a majority number (quorum). With only 2 nodes, there is no majority and therefore at least 3 are required for redundancy (think of RAID with Parity disks). The controller is the critical piece of the software. Here all ARP requests are handled and all information is shared with the ESXi hosts. The Controller will also take care of multicast replication when the unicast mode is chosen for transport.

Prepare hosts

The following step is to install 3 kernel modules in all hosts. This is done using the VMware Update Manager (the Windows application) and just a single click in the GUI. It will install the NSX module to communicate with the controllers, the Distributed Logical Router (DLR) and the Distributed Firewall. No configuration is necessary as all changes are pushed by the controllers.

NSX Edge

To get traffic to and from the virtual environment you will require an NSX Edge today, as soon as OVSDB is available this could be your MX router as well! The NSX Edge has 2 forms, the first is the Distributed Logical Router that offers basic routing, bridging and firewall features. The second version is a VM that traffic is hair pinned through, like a software router. The second version also offers VPN and Load Balancing features.


To support the VXLAN transport, the hosts need to get an IP address to use as source address. This can be assigned using DHCP or an IP Pool, more on this in a next blog! Next a Segment ID pool needs to be created to allocate VNI’s for the virtual networks (called logical switches). Today it’s not supported to have more than 4000 segments on NSX, which combines the regular port groups and NSX port groups on the distributed vSwitch.

Transport Zone

The final step is to create a transport zone. This defines a “data center” or a “site” where NSX runs. This limits the broadcast domain. Per transport zone you are able to select wether the NSX Controller should handle multicast replication (Unicast mode) or that the network should handle this (Multicast or Hybrid mode). The recommended choice is Unicast mode so you won’t require multicast routing protocols or IGMP from the network. VMware eliminates the need for smart network hardware even further with this, but the solutions runs best with optimized hardware of course running as IP CLOS Fabric offering “Layer 2 as a service” using network virtualization.

Logical Switches

Now the deployment is ready to have the first VXLAN segments created, which are called Logical Switches. These networks are created as distributed port groups, so you can use the standard tools to connect VM’s to these virtual networks and suddenly you are using VXLAN transport!


I hope this article gave you some insight in how NSX is configured and what components it consists of.


Happy Labbing!


Rick Mur

EVPN Configuration

In my last blog I explained the features and use cases of the EVPN technology. In this blog I want to show how easy it is to configure, enable and expand EVPN. The configuration is focused on the Juniper MX platform, but as Junos is the single operating system across the entire Juniper portfolio, configuration on other platforms (like EX9200) is equal.


The topology is really simple. I’m using 2 routers in this example, so multi-homing is not in scope. Each router has an Ethernet segment connected that consists of multiple VLANs on each side. There is 1 VLAN ID that is not equal on both sides, so this has to be taken care of.


Prepare for EVPN

To make sure we can start creating our VPN, we have to ensure the foundation is in place. This means we need IP reachability to the other Data Center routers loopback address and we need BGP with the EVPN address family enabled. The Junos release I’m testing with needs a special knob to be enabled to ensure packet lookups are done in the right way. Other versions will see this knob disappear as it is default moving forward.

Interface ge-0/0/1 is the one connected to the MPLS network and we configured an iBGP neighbor for both the Layer 3 VPN and EVPN address families to exchange layer 2 and layer 3 routes with each other.

We use OSPF as internal routing protocol, but in this simple topology it could’ve been a static route as well.

VLAN Based Service

There are 2 options to configure EVPN. The first option is to configure a separate EVPN Instance or EVI for each VLAN/Bridge-Domain. This brings the best separation of traffic, MAC address advertisements and give full control over flooding of broadcast traffic on a per VLAN/Bridge-Domain basis. The VLAN ID of traffic that is sent between the 2 PE routers will be set to 0.

EVPNoption1A Route-Target and Route Distinguisher is created for each EVI and therefore each VLAN has it’s own MPLS Label.

This could potentially cause scaling issues, as PE routers have a limit on the number of VPNs that can be configured.

The following is an example of the configuration for our topology.

The commands “flexible-ethernet-services” and “flexible-vlan-tagging” on the interface allow us to use this interface for a variety of services. This means we can use single or double-tagged interfaces and we can configure both layer 2 bridging as layer 3 sub-interfaces on  this single physical interface. One of the very strong features of the MX platform!

As you  can see the VPN configuration is very easy! We just make sure the right interface is in the VPN and we assign a RD and RT value to it and we’re done!

As we have a single-homed scenario, there is no configuration necessary. The ESI (as discussed in my previous blog) is by default set to 0 on single-homed routers and doesn’t have to be different on the other end, because of the single-home setting, which is enabled by default.

VLAN Aware Service

The second option is more flexible to use. This is the VLAN Aware service. This option allows multiple VLANs and Bridge-Domains to cross a single EVPN Instance (EVI). This improves scaling and still allows for proper separation. The VLAN ID is now encapsulated in the packet. To ensure proper ECMP forwarding in the MPLS network an MPLS label is assigned per VLAN ID, so that traffic will be load balanced per VLAN and not per EVI.

EVPNoption2The configuration for this service is a bit more complicated, but still fairly simple and also still allows for VLAN translation! Even MAC address overlap is allowed within this service between different VLANs/Bridge-Domains.

As you can see we also ensure the VLAN translation is taken care of. We could use VLAN normalization in this case, which means we use a different “unique” VLAN ID within the EVPN service and use VLAN translation on each side to ensure the local side has the right encapsulation.


This blog showed you how easy and how flexible EVPN configuration on Junos is and how fast we can set-up a optimal Layer 2 Data Center Interconnect. In the following blogs I will discuss multi-homing and Layer 3 integration to ensure all-active routing across Data Centers.

EVPN (RFC 7432) Explained

EVPN or Ethernet VPN is a new standard that has finally been given an RFC number. Many vendors are already working on implementing this standard since the early draft versions and even before that Juniper already used the same technology in it’s Qfabric product. RFC 7432 was previously known as: draft-ietf-l2vpn-evpn.

The day I started at Juniper I saw the power of the EVPN technology which was already released in the MX and EX9200 product lines. I enabled the first customers in my region (Netherlands) to use it in their production environment.

EVPN is initially targeted as Data Center Interconnect technology, but is now deployed within Data Center Fabric networks as well to use within a DC. In this blog I will explain why to use it, how the features work and finally which Juniper products support it.


Data Center interconnects have historically been difficult to create, because of the nature of Layer 2 traffic and the limited capabilities to control and steer the traffic. When I have to interconnect a Data Center today I have a few options that often don’t scale well or are proprietary. Some examples:

  • Dark Fiber
  • xWDM circuit
  • L2 service from a Service Provider
  • VPLS
  • G.8032
  • Cisco OTV or other proprietary solutions

Most options also only work well in a point to point configuration (2 data centers) like Dark Fiber or xWDM circuits or only with a few locations/sites (VPLS).

Only the proprietary solutions have a control-plane that controls the learning and distribution of MAC addresses. All others are technically just really long Ethernet cables that interconnect multiple Data Centers.

Now why is that so dangerous? 

I’ve had numerous customers with complete Data Center outages because of this fact. No matter which solution you pick that does not offer a control-plane for the MAC learning, when problems occur in one DC like ARP flooding or other traffic floods. They will propagated to the other DC as well using the interconnection layer that can only protect it with features like storm-control, etc. This means that impact on Layer 2 in DC1 impacts other Data Centers. Where in most cases the other Data Center is in place because of high availability reasons.

EVPN will solve this!


EVPN is technically just another address family in Multi Protocol (MP) BGP. This new address family allows MAC addresses to be treated as routes in the BGP table. The entry can contain just a MAC address or an IP address + MAC address (ARP entry). This can all be combined with or without a VLAN tag as well. The format of this advertisement is shown in the drawing.


BGP benefits

Now the immediate benefit of using BGP for this case, is that we know how well the BGP protocol scales. We won’t have any problems learning hundreds of thousands of MAC entries in the BGP table.

Route L2 traffic – All-Active Forwarding

The second benefit is that we now have MAC addresses as routing entries in our routing table and we can make forwarding decisions based on that. This means we can use multiple active paths between data centers and don’t have to block all but 1 link (which is the case in all the previously mentioned DCI technologies). The only traffic that is limited to 1 link is so-called BUM traffic (Broadcast, Unknown unicast and Multicast). For BUM traffic a Designated Forwarder (DF) is assigned per EVPN instance. This technology is not new and is also found in protocols related to TRILL and other proprietary technologies.

ARP and Unknown Unicast

ARP traffic will be treated in a different way than regular Layer 2 interconnects will do.  First of all Proxy-ARP is enabled when EVPN is enabled on an interface. The Edge gateway of the DC will respond to all ARP requests when it knows the answer to them. This immediately implies a very important rule. When an ARP request is done to an IP address the Edge router doesn’t know or when traffic is received with a Destination MAC that is unknown, the Edge gateway will drop this traffic! This limits unknown unicast flooding immediately to only a single DC, which is quite often a cause of issues. The Edge Gateway has to learn about a MAC or ARP entry from the Edge gateway in the other DC before it will allow traffic to pass.


We obviously want 2 gateway devices at the edge of our Data Center. Therefore we have to support a multi-homing scenario. EVPN allows for an all-active multi-homing scenario as previously explained, because of the routing nature. Besides this we need to limit the flooding of traffic, which is the reason why we choose a Designated Forwarder per instance which is responsible for receiving and sending BUM traffic. We also support split-horizon, to prevent that traffic originating from our own Data Center does not get back in through the other Edge Gateway. This is done using the ESI field in the advertisement. An ESI is an identification for a certain “Data Center Site”. Both gateways in a single Data Center should use the same ESI number, to prevent traffic getting looped back in. Other Data Centers should use a different ESI number. This means that each Data Center will have 1 ESI number.

MAC Mobility

The EVPN feature is designed for extremely fast convergence. Another big issue within Data Center is what’s called MAC flapping and mobility. This means that when a Virtual Machine moves with it’s MAC address, the MAC needs to be re-learned within the network and the old entry should be deleted. Within a single layer 2 switch domain this happens quite fast, which is also the cause of issues, when a duplicate MAC address is found in the network. Then the same entry is learnt on two different ports causing “MAC Flapping”. EVPN solves this by introducing a counter when a MAC moves between Data Center locations. When this happens too often in a certain time period (default 5 times in 180 seconds) the MAC will be suppressed until a retry timer expires to try again. The benefit of the MAC sequence numbering is that when the Edge Gateways see an advertisement for a MAC address with a higher sequence number they will immediately withdraw the older entry, which benefits convergence time after a VM move.

Layer 3 All-Active Gateway

The last benefit I want to highlight is that, because the Edge Gateway is fully aware where hosts are located in the network, it can make the best Layer 3 forwarding decision as well. This means that Default Gateways are active on all edge gateways in all data centers. The routing decision can be taken on Layer 2 or on Layer 3, because the edge gateway is fully aware of ARP entries .

 Data Planes

Currently there is 1 standard and 2 proposals that all rely on the same control-plane technology (EVPN). The first is the current standard RFC7432 which is EVPN with MPLS as it’s data-plane. The second proposal is EVPN-VXLAN, where the same control-plane is used, but the data-plane is now either VXLAN or another Overlay technology. The third proposal is to use PBB encapsulation over an EVPN control-plane with MPLS as data-plane.



Finally I want to discuss which products support this technology and how you can implement it. Juniper currently implements RFC 7432 (EVPN-MPLS) on it’s MX product line, ranging from MX5 to MX2020 and on the EX9200 series switches. The EVPN-MPLS feature is designed to be used as Data Center Interconnect.

Since the MX has full programmable chips Juniper also implements several overlay technologies like VXLAN. This means we can stitch a VXLAN network, Ethernet bridge domains, L2 pseudo wires and L3 VPN (IRB interface) all together in a single EVPN instance.


All this means we have a swiss army knife full of tools to use within our Data Center and we can interconnect various network overlays and multiple data centers together using a single abstraction layer called: EVPN!

In my next blog I will demonstrate the use of EVPN on the MX platform and how easy it is to implement an optimal Data Center Interconnect!

Nested Virtualization

A typical Network Virtualization demo is difficult as you need quite some hypervisor hosts to run some VMs on and interconnect them using Overlays. I solve this using nested virtualization. This means that I run a hypervisor running on another. This gives me the flexibility that my physical nodes, or “hypervisor underlay” if you will, can scale easily and I’m independent of them.

My physical cluster consists of 2 nodes running ESXi with vCenter. On top of that I’m running 4 other ESXi hosts divided in 2 “virtual” clusters and 4 KVM hosts as Contrail Compute Nodes.

How does this work?

This technology works using Intel’s VT-x (which is hardware assisted virtualization) and EPT (to virtualise memory allocations). This combination works since the “Nehalem” archnested1itecture (released 2008). The technology is ported to the more “Desktop” oriented CPU’s as well, so there is a good chance your notebook supports it as well. Since the Haswell architecture the nested virtualization works even better as Intel now supports VMCS Shadowing for nested VMs, which creates a data structure in memory per VM (and now supports nested VMs as well, which used to be a software effort).

Memory is the biggest burden in these nested set-ups. CPU performance is always hardware assisted, so nested VMs almost feel the same as regular VMs. The problem is that ESXi itself requires 4GB of memory to run, it requires around 2GB to run properly. So get at least a 32GB server, otherwise you run out of memory very fast!

The amount of nested virtualization is technically unlimited. So you could have 2, 3 or 4 levels deep. Where at some point to use-case becomes very small of course. The same goes for 32-bit and 64-bit machines. As long as your physical CPU is a 64-bit one (32-bit CPU’s do not support VT-X or EPT) you are able to run 32-bit and 64-bit guests inside a nested ESXi installation running on top of ESXi.



I’m using ESXi as my main hypervisor as it has the best management  tools currently out and most of my customers are using it. This gives me a stable foundation. So this set-up is based on the ESXi version of configuration. I’m using the latest version of ESXi as of this writing (5.5u2). To set-up nested virtualisation properly you have to use at least Virtual Machine hardware version 9 or 10. This means you have to configure this using the vSphere Web Client and therefore vCenter Server (Appliance) is also required to be running in your lab. I would also recommend to give your virtual ESXi at least 2 vCPU’s (no performance difference in allocating cores or sockets) and sufficient memory. Remember that ESXi will consume about 2GB to run for itself and the rest you can allocate to VM’s. Depending on the amount of nested VM’s you want to run I gave 6GB of memory to my ESXi as I’m only running 3-4 guest VM’s on this nested hypervisor.


The most important setting is when you open up the “CPU” panel in VM settings. There you have to enable “Expose hardware assisted virtualisation to the guest OS”. This will enable the exposure of the Intel VT-x feature towards the VM and will enable ESXi to recognize a virtualization capable host. If this checkbox is greyed out, it means your CPU is not suitable for nested virtualization.

(The same steps are required to install nested KVM inside ESXi, Hyper-V may require some additional tweaking)



After installing ESXi inside the VM you are almost up and running! First of all you probably want to allow VLAN tagging to go towards the nested ESXi and for VLAN tags to come from the vSwitch running inside the nested ESXi. This requires a port group that is set for VLAN Trunking. I’m using a Standard vSwitch, not a Distributed vSwitch (DVS) on my physical cluster, because of the fact that the vCenter server has to be up and running to be able to power on a VM on a DVS. Because I will be playing around with this environment I can’t rely on vCenter being up all the time. So therefore I chose to use Standard vSwitches on the physical cluster. As I will experiment with networking features on the nested ESXi installations. To create a VLAN trunking capable Port Group, you have to create this on each physical host separately and allow a VLAN tag of “all” or “4095”.


This will make sure that my network inside the nested ESXi installation is VLAN tagging capable!

Now before this will work fine, we need to change one more setting on our physical cluster ESXi installation.ESXi by default considers itself to be the master of the network (at least inside the ESXi environment). Because ESXi is generating the MAC addresses that the VMs use it doesn’t allow any other MAC addresses to enter the vSwitch coming from a virtual interface. This is a great security feature and definitely helps a lot in preventing all kinds of network attacks to be based from inside the VM. Because we will be having a whole new ESXi installation that is generating it’s own MAC addresses we do want to allow unknown MAC addresses to enter on the virtual interfaces connecting to the nested ESXi VMs. To enable this you have to allow the “Promiscuous mode” on the vSwitch or on the port groups you applied to the ESXi VM. I enabled it on the vSwitch itself, just to be sure not having to worry about it when I create a new port group. Remember to change this setting on all of your physical hosts when using a Standard vSwitch. Make sure to change “Promiscuous mode” to “accept” and that “Forged transmits” and “MAC address changes” are also set to “accept”, but that’s the default.


Now Promiscuous mode does impact performance on your ESXi server. ESXi has no concept of MAC learning like a regular Ethernet Switch. This means that when we disable the security control of MAC addresses being assigned to the virtual interfaces (promiscuous mode) we will flood all traffic to all of the ports inside the vSwitch. With regular feature testing you won’t generate much data traffic, but when you do get to a few Mbps of traffic, this could impact CPU performance quite a lot. There is a plug-in that will enable MAC learning on interfaces that you configure it for. This does require a Distributed Virtual Switch to be running in the physical cluster. You can find it, together with implementation details, here: https://labs.vmware.com/flings/esxi-mac-learning-dvfilter

I would not recommend using it in tests with a lot of NFV appliances (virtual firewalls, virtual routers, etc.), as the entries learnt through this filter will not age out. This means that when your nested Guest VMs are moving a lot, it will register that MAC address on each port meaning the performance improvements will go away. If you have a stable environment then this MAC learning feature will be beneficial.  More details are found in this excellent blog: http://www.virtuallyghetto.com/2014/08/new-vmware-fling-to-improve-networkcpu-performance-when-using-promiscuous-mode-for-nested-esxi.html


The final optimization that can be done when you are running ESXi as a VM is to run VMware Tools. This will make sure that a graceful shutdown can be done and makes your life a bit easier when it comes to IP addressing and have visibility inside vCenter. The default VMware tools do not support ESXi, you require to have the version installed found here: https://labs.vmware.com/flings/vmware-tools-for-nested-esxi


When everything is installed I highly recommend creating a separate cluster for these nested hypervisors and keep your physical cluster clean. Now you are ready to deploy VM’s just as you are used to, but now inside a safe environment where you can start breaking things!


Happy Labbing!

« Older posts

© 2017 Rick Mur

Theme by Anders NorenUp ↑