ECMP next hop on Juniper M T and SRX series routers

If like me, you have to jump around customer requirements, you may one day find yourself in a situation where you need to utilise capacity on 2 or more links between locations. My preference is to bond my uplinks with 802.1ax/802.3ad/LACP and let the upstream provider deal with the rest. Sometimes the providers let you down and can do nothing. They cant run LACP from their edge device to you, and they can’t transit your LACP frames so that you can perform your own LACP between locations. Also sometimes you have multiple links for different providers.

In this situation your last resort is Equal Cost Multi Path(ECMP) next hop. If you have a 2 or more routes in your routing table with exactly the same metrics and there are none that are more preferred, an ECMP decision is triggered. On Juniper routing platforms this is quite rudimentary in that one of the routes will be chosen (at random or based on src/dst hashes) for a particular route and installed in the FIB (the hardware forwarding engine). This means that the effectiveness of the traffic spread is limited to the number of routes in your table in a particular direction.

My typical implementation involves running OSPF between routers on each link with identical metrics. From the “remote” end of the network I do not aggregate the advertised prefixes, as this would reduce the pool of routes, and instead advertise all prefixes individually. This is often a whole bunch of /32 point to point customer IP’s and this is also partially why I choose to use OSPF for this.

Advertising from the core however is a bit more of a problem. Here typically you are advertising mainly the default route. There may be some peering routes that you have on either end and you may include those too, but typically you do not want to be sending a full table to some remote end of the network, as usually the reason you are here in the first place is that you are resource constrained.

The practical upshot is that traffic will balance ok in the direction towards the “remote” node, but very little or not at all inbound from the “remote” node. Typically this is the “download” direction and usually the direction most of the load is in in any case, but our situation is not ideal.

To achieve a better spread, and to not have to worry to much about how many routes you are using, you need to implement a policy on the forwarding table. I know it sounds like I made that up, but yes, thats a real thing. If you do not do this then your traffic spread/diversity will be constrained by the points discussed above.

So we create the policy..

set policy-options policy-statement my-default-balancing-policy then load-balance consistent-hash

And then apply it to the forwarding table..

set routing-options forwarding-table export my-default-balancing-policy

This will now let your traffic use all equal routes instead of just the selected one.

Your 2 balancing options are consistent-hash and per-packet. Per packet will send packets down each link in a round robin fashion and will result in nearly perfect load spread. However, this will cause out of order packet delivery between the sites as there will always be performance differences on the links which is why I never use it. The performance impact of out of order packets, on TCP specifically, is significant. The consistent-hash looks at the traffic IP source, destination and protocol fields and uses those values to calculate which link to use. This is good at keeping traffic flows on one path and packet delivery consistent.

ECMP algorithm choice on the MX series platform is performed quite differently, but many of the points discussed above are still valid. This is to be expected as the MX is a routing and switching platform so hashing at multiple layers is possible (L2/L3/L4) There are many more options to consider and we will leave that for another time.

A final note, the above hash looks at L3 information as a key for hashing and on an MPLS enabled network this may not be enough. You can also set ECMP options for MPLS with the following statement.

set chassis maximum-ecmp 16

Options are 16/32/64 and allow for up to that many alternate LSP to load balance across (thats if you have multiple LSP’s to your destination).

Juniper M10 value propositon

The Juniper M series of routers have been obsoleted but are a really good value proposition if all you need is a few gig of reliable routing capability. The M10/M10i is a redundant H/A solution that is readily available in the refurbished market from sub 2000 USD and the most cost effective way to bootstrap a small enterprise with a robust core.

What You Get

  • 5U Chassis
  • 2 x C-FEB Forwarding Boards (active + standby)
  • 8 x PIC Physical Interface Card slots
  • 2 x Routing Engines 400Mhz CPU 256M RAM (higher spec available for more $)
  • 4 x 300W PSU (2 Required for operation 3/4 for redundancy)
  • In Service replaceable fan tray

Pro’s

  • Fully redundant PSU, routing engine and switch fabric.
  • High availability features
  • Cheap and available
  • Enterprise grade
  • MPLS and IPV6 support
  • Dedicated out of band ethernet and RS232 management ports
  • Plenty of SONET/SDH/ATM PIC options at reasonable pricing

Con’s

  • Limited capacity (1G per PIC/slot)
  • Relatively inefficient (power and size vs throughput)
  • End Of Life
  • No layer-2 capability (kind of.. see below)
  • Only vlan-tagging support, no stacked-vlan-tagging or flexible-vlan-tagging (ie. no L3 support for  Q-in-Q).

The 1G limitation per port is the hard limit on this device and I would not want to try and use it anywhere beyond a total of 4G of capacity as balancing evenly across ports starts to become a factor. The M series routers are IP/MPLS routers supporting all the standard BGP/IS-IS/OSPF and MPLS protocols as well as allowing for multiple routing instances.

Stateful firewalling and VPN is possible but would require a services PIC, two for redundancy. If this is what you are needing then you are generally better off looking at something else as they can be expensive and limited to 1G per module.

Now the M series were made in an era when layer-2 and layer-3 functionality was typically serviced by separate devices, so the M series routers are just that, routers, they have no switching capability at all. Well kind of, it support 802.3ad LACP link bonding between PIC’s and with the progression of technology and standards, Junos and the M platform received upgrades and features which included MPLS and VPLS functionality which is technically a layer-2 technology.

Because of the lack of switching support VPLS is limited in what it can do and you can run into issues if you are not aware of of these limitation. How we typically implement is as follows

  • 4 x 1G ethernet PIC’s
  • 2 x 2 Ports bonded with LACP into 2 aggregated ethernet ports on switches

(these numbers can be doubled for more capacity)

We use the 2 aggregated ports to provide a network facing port and a services facing port. They also provide link redundancy into the network, they typically all uplink into the same switch stack and are used to provide a pair of interfaces/VLAN’s for the M10 which can also be used to loop traffic where required. You could do this all with one AE port with all the physicals in it and just use VLANs but I like to split my roles across interfaces for easier visibility and troubleshooting.

So on one VLAN on the network AE we set up MPLS capability with all the layer-3 stuff that’s required to make MPLS work. On the services port we setup a customer or service facing VLAN that we want to tunnel using VPLS. This is done by setting the VLAN port encapsulation to vlan-vpls and then creating a routing instance of  instance-type vpls and adding the interface to that instance type. Now this is technically a switching function that is being performed on a routing only device.

The caveat is that you need to create a separate routing instance for every layer-2 service that you want to use. You CANNOT use VLAN’s on the VPLS service because you will run into MAC learning issues due to the fact that the M10 is not layer-2 aware and cannot differentiate between the different broadcast domains of multiple VLAN’s. It will work for a bit but you will run into random dropped packets as the MAC learning table on your endpoint devices gets polluted.

The flexible-vlan-tagging or stacked-vlan-tagging option on interfaces is allowed but ultimately not supported. On commit the device spits warnings in the messages log and when you try and configure the inner and outer tag the router will not accept the configuration. You should configure the vlan-tagging option instead.

A simpler supported L2 feature is the l2circuit using MPLS.  It is a point to point only tunnel that does not perform any MAC learning whatsoever, it just take the frame on one side and spits it out on the other. This can be configured on VLANs on ethernet ports if the encapsulation type on the VLAN is set to vlan-ccc. The port will accept further tags if they are present as well as “L2 local” frames such as LACP,LLDP,STP BPDU’s etc… The service is only really limited by overall MTU. This is because the M10 is not involved in any L2 learning so it will transparently pass the frame from one endpoint to the other endpoint. This is also why you can only have one endpoint because the M10 cannot make a path determination with no address information.

The down side is that troubleshooting can be a bit harder in that you cannot see any learned MAC’s but the up side is that you do not need to worry about memory + MAC learning limits.

As mentioned above, the lack of L2 support means that we usually pair an M10 with 2 EX4200’s in VC mode. QFX would be better but we are looking at a budget solution here so they don’t make sense. This gives you a certain amount of L2 flexibility that will cover most use cases. Be aware that EX series switches only support VLAN swap and push functions NOT VLAN pop. This can be somewhat limiting in this environment. One final note regarding the EX configuration for l2circuits, you can configure “dot1-tunneling layer2-protocol-tunneling all” on the EX4200’s which will ensure you can transparently take all frames from a customer facing VLAN to a l2circuit on the M10. This is also where we can look at MAC learning for troubleshooting as the switch will learn customer MAC’s, and also where we set MAC learning limits to prevent possible issues introduced on the EX by customer networks.

Saving IP space on your point to point peering subnets using /31 subnets aka RFC3021

Most of us are used to allocating a /30 subnet for point to point peering, which gives us 2 device IP’s, a network address, and a broadcast address. With the exhaustion of IPV4 space it’s useful to be able to save wherever you can, so where possible you should consider /31 allocation instead of /30. Most modern day vendors support RFC3021 and using it is as simple as just specifying the /31 subnet mask. I have used this on various Juniper and Cisco devices in the past without any issues, but unfortunately not all vendors or models support it so your mileage will vary.

I would also, always recommend performing further testing beyond just a ICMP reachability tests if broadcast protocols are involved. A prime example is if you use something like the OSPF routing protocol that uses broadcast as a discovery mechanism. It should work but if it does not then possibly see if you can set it o P2P operation before reverting back to a /30 allocation.
I need to mention Mikrotik specifically here because they do not support RFC3021 in an obvious way. If you just specify x.x.x.x/31 it wont work. You need to specify a /32 with a matching network address. So 1.1.1.0/31 would look like so

/ip address add address=1.1.1.0/32 network=1.1.1.1 interface=ether1

The above example is compatible with RFC3021 but is actually different in that the network address does not have to fall within the same /31 as the IP. I have struggled to find the RFC for this assignment scheme but it allows you to be able to add the same IP multiple times with a different network address. This allows you to reach multiple devices on the same network segment using only a single IP per device. This obviously only works if all devices support this format.

Example of using this to connect to 2 other routers with IP’s 2.2.2.2 and 3.3.3.3.

/ip address add address=1.1.1.0/32 network=2.2.2.2 interface=ether1
/ip address add address=1.1.1.0/32 network=3.3.3.3 interface=ether1

Making your Debian server networking redundant

You will need at least the following…

  1. A pair of stacked switches that support creating an LACP bonded port across the stack on 2 different nodes. This gives you the best of all worlds being able to provide redundancy and increase your bandwidth.
  2. Or alternatively, 2 ports on the same or on different unstacked switches. This is the bare minimum you can do to mitigate link failure. Note this setup has no polling mechanism so if the physical ethernet link stays up but is not operational because of device switching failure, or a failure on the another port on the device that provides the uplink, then this wont help you.

On your server you will need 2 (or more) network cards and some “simple” setup

Install the packages that you will need in case you don’t have them already.

  • apt-get install ifenslave vlan bridge-utils

The example sets up the following

  • eth0 and eth1 bonded together into bond0
  • create 2 bridges br8 and br9
  • create 2 vlans bond0.8 and bond0.9
  • place them in each bridge respecitvely
  • add IP details on br9
  • br8 has no L3 config on it and in this specific case is used by KVM to bridge virtual machines into as they come online

For option 1 edit your /etc/network/interfaces to look something like this


# This file describes the network interfaces available on your system
# and how to activate them. For more information, see interfaces(5).

source /etc/network/interfaces.d/*

# The loopback network interface
auto lo bond0 bond0.8 bond0.9 br8 br9
iface lo inet loopback

iface bond0 inet manual
 bond-slaves eth0 eth1
 bond-mode 802.3ad
 mond-miimon 100
 bond-use-carrier 1
 bond-lacp-rate 1
 bond-min-links 1
 # send traffic over the available links based on src/dst MAC address
 bond-xmit-hash-policy layer2
 mtu 1600

iface bond0.8 inet manual
iface bond0.9 inet manual

iface br8 inet manual
 bridge_stp off
 bridge_ports bond0.8

iface br9 inet static
 address 192.168.0.2
 netmask 255.255.255.0
 gateway 192.168.0.1
 bridge_ports bond0.9
 bridge_stp off

For option 2 edit your /etc/network/interfaces to look something like this (only the bond0) config changes


# This file describes the network interfaces available on your system
# and how to activate them. For more information, see interfaces(5).

source /etc/network/interfaces.d/*

# The loopback network interface
auto lo bond0 bond0.8 bond0.9 br8 br9
iface lo inet loopback

iface bond0 inet manual
 slaves eth0 eth1
 bond_mode active-backup
 bond_miimon 100
 bond_downdelay 200
 bond_updelay 200

iface bond0.8 inet manual
iface bond0.9 inet manual

iface br8 inet manual
 bridge_stp off
 bridge_ports bond0.8

iface br9 inet static
 address 192.168.0.2
 netmask 255.255.255.0
 gateway 192.168.0.1
 bridge_ports bond0.9
 bridge_stp off

Most use cases probably will not require bridging or VLAN but I thought it best to provide examples of the entire feature set, you can always reduce to what you need.