I love how this starts off "What is a VXLAN..." that was going to be my first question! So often posts like this seem to assume everyone knows what the topics are at the start. I know what WireGuard is, and I've at least heard of VXLAN before, but I couldn't remember what it was.
You may consider VoIP phones. When a phone boots up on a network segment the DHCP process comes into play. This process is a Layer 2 process. The DHCP packet could contain a boot server field (this is typical) so the VoIP phone can grab configuration.
One may want the boot server info "isolated" from other network segments. Utilizing VLANs is one way to do this. Additionally, it is typical that QoS is applied at Layer 2 (something necessary for real-time protocols like VoIP).
Over VXLAN yeah... You won't believe how many small-medium companies are moving from on-prem to small managed private clouds. VXLAN allows them to maintain their existing network configuration. "You hope" ha! I suppose by this you mean that the wireguard + vxlan may not be a "mature" combination for something as mission-critical as voice traffic? VXLAN implementations by mature companies (Netgate, Fortinet, etc...) those work. Maybe this wireguard+vxlan would work too.. who knows!?
Edit: I see you changed your comment to consider remote DHCP which causes my comment above to be irrelevant. Oh well. The fact still holds that VXLAN is meant to handle Layer 2 so saying BGP is an alternative is like putting a square peg in a round hole.
I appreciate that you are trying to give examples of "L2 things" but none of this is really good advice.
For QoS, what do you think double-encapsulating all of the QoS bits inside of two tunnels is going to do for you? If you pay your carrier to respect QoS, you have to actually give them the bits! QoS goes on the OUTSIDE of the onion.
DHCP for VLANs should be handled by snooping for it on the switch and forwarding over L3 (ip helper-address), not by running vxlan over a VPN tunnel. And really it shouldn't ever be done over a potentially unreliable WAN link either; you should have a local agent for DDI at each site.
The only story here is that L2-over-L3 VPN have their very specific niches, and VXLAN+Wireguard is an alternative to L2TP+IPSec that is (IMO) both easier to configure and more reliable when such a thing is needed.
Sorry for the dangling preposition. "That" refers to VXLAN+Wireguard being easier and more reliable. It has the drawback though of having very high overhead at 130 bytes/packet, and it can be very tricky to use over the public Internet without paying lots of special attention to tuning the MTU of all devices on the bridged segment. I'm glad to know that there are geeks who have not been exposed to the horrors of IPSec, though. And yeah, you should pretty much never do L2VPN. It is perilous in the extreme.
I've found that they work just fine. In my case all that was required was setting a DHCP option pointing them at their controller. The QoS to make it work well under load would've been the same for a L2 tunnel.
As someone that has fight for years with L2 in the WAN, I will not advice that at all.
Just installa a DHCP server and transport the traffic over L3. You have lot more control both on routing both on QoS
But doesn't provide layer 2 between networks. Think of devices that are hardcoded to communicate with broadcast or multicast with TTL of 1, you either need some active reflector, and cope with any perculiarities of the device, or you simply extend a single vlan between two routers (using vxlan or another solution)
I sometimes need to extend a system like this from one site to another. One is a calrec system (an audio mixer, I think it's the control traffic that needs to be sent), and I don't have enough access or time to see if I could build some kind of transparent proxy -- it won't work with multicast routing.
I do however have enough time to create a layer2 network between two nics. I tend to use mikrotiks for that, create an eoip tunnel (GRE with proprietary addons to cope with fragmentation) between the two endpoints and pop the interface in a bridge with a physical port, and move on.
I've been in or very close to networking for a long time. Here's a certainty: there's lots of ways to do anything. Here's a corollary to that: someone will be sure to tell you how bad any choice you make is.
The feature I would be interested in is if this can do link state toggling on the on the vxlan interface if the wg handshake timer goes stale. If that works, then it becomes practical to do things like run ospf routing over the vxlan interface.
What exactly would you be intending to accomplish? OSPF already has state timers, and furthermore runs just fine on a wireguard interface without having to introduce a vxlan tunnel.
There's a weird font-rendering bug on this site that causes the text in the code blocks to be unreadable unless you highlight it with your mouse. If you enable Javascript, it seems to fix it.
Not sure if the author is reading this thread, but it's something you may find worth investigating fixing.
IP over Ethernet over VXLAN over UDP over IP over WireGuard over UDP over IP over Ethernet… sigh
OpenBSD does support both routing domains and multiple routing tables and includes multiple routing daemons in the base system. I would recommend to the author to stop hacking at the keyboard, grab whatever not to structured visualisation tool works for them (e.g. a whiteboard, a block of paper, a random drawing app, Visio) and (re-)phrase the problem. Are you solving a problem or showing of how many acronyms you can expand without looking them up? This n layer encapsulation can work and can even be required to reproduce some (problematic) organisational structure, but it's far from elegant. Given the chance I would vastly prefer to just use multiple routing domains for the WireGuard tunnel interfaces and the underlay. It would result in far less complexity to manage as well as less overhead.
Why do so many people insist on tunneling Ethernet over IP? What's keeping operators from using IP routing (and just one layer of encapsulation) instead? Is IP routing so scary or everyone that indispensable applications that only work over Ethernet?
Well an L3 network is routed, so you just have routes. 32 bit route for the "server", whatever mask to reach others through the "server". The server has 32 bit routes to each client, so it works.
This is essentially how you want to build your DC networks today too (why waste ips on netid and broadcast that you never use) (for public addresses).
But you stick BGP on there to exchange route information rather than static routing.
Right. My thinking is that with VxLAN you can get each tenant in the datacenter to have its own L3 network, isolated from other tenants L3 network.
Each tenant can have multiple subnets; run BGP, etc. and the IP addresses can be re-used between different tenants.
It doesn't seem wireguard can do that with just p2p network. This was one argument I had for adding VXLAN encapsulation over IP for the above comment. You also get L2 connectivity if you want it.
Ah, yes. Well you can set up multiple tunnels for lower overhead too, all you need is an interface so you can bind it to a VRF. You save 50 bytes on every packet though in reality either way works.
As long as you're lowering MTU and not bridging your VXLAN you're fine.
You can run VXLAN over any L3 so it's versatile like that, but the benefit of VXLAN over other encap protocols is that it's simple, accelerated in a lot of hardware, integrated with BGP-EVPN to make scaling easier, more predictable and traceable.
I end up having to run basically this very setup (on OpenBSD, too) because I have a customer who has a Novell NetWare 5 setup and runs IPX only. Bad times.
I used to run a cluster of DNS/DHCP servers that were the first on site to run 5. The rest were 4.11 until we binned them for 6. Three cream coloured Compaq 3U lumps.
Right, but they weren't running IP and refused to do it. I did set up a /30 so the poor old thing could synchronize its clock via NTP, but that was the only IP it talked.
This one's running (present tense) on a 1 GHz Socket 370 Pentium 3. It's got some weirdness about Pentium 4 and newer CPUs. I think it can be patched up but the client doesn't want to pay for the work because "it's fine like it is." Not worth the headache to virtualize as the first attempt didn't work.
vMotion needs L2 adjacency to make live migrating VMs easy. Some software rely heavily on broadcast discovery messages and are thus designed for LAN usage not Internet connectivity but businesses try to stuff a square peg into a round hold.
Just the grateful that nobody has tried to wrap the entire thing in JSON over HTTP yet! I wouldn't be surprised if we get Wireguard over websockets for "enterprise" applications soon.
Sometimes you just need an L2 tunnel. Most of the time you don't, but when you do, you do. For example, if you use IPv6 over SLAAC in a private network, you'll need to route NDP.
In the rare cases that you do need an L2 tunnel between two different locations, you probably want some kind of authorisation and authentication of the traffic to prevent injection/spoofing attacks and to make life just a but harder for the NSA (Google's use of HTTP was one way the NSA managed to tap connections that were otherwise encrypted by HTTPS). After all, this isn't just any traffic, these are internal subnets.
In terms of authorised traffic, Wireguard is quite lightweight and foolproof. Perhaps IPSec is even more lightweight but it's a pain to set up. The alternative would be to wrap all internal network traffic in an encrypted protocol and set up the necessary whitelists in the upstream ISPs.
The impact of such layering depends on the network connection between the data centers. If you can get jumbo packets across, fragmentation won't be a problem at all. If you run your own fiber between data centers, there's basically no downside until you're reaching very high saturation network saturation.
This can cause massive packet fragmentation. I'd be most interested in the performance degradation due to the L2 encapsulation. Are there any benchmarks available for this kind of project?
That's what I was thinking, unless you have jumbo frames you're going to have a hard time stuffing ethernet frames into IP payloads. Does Vxlan mitigate this somehow?
I've tunneled VXLAN over Wireguard on Linux. In my setup, my WAN's MTU was 1500 bytes, and my Wireguard tunnel's MTU was 1550, with the VXLAN's MTU being 1500. Surprisingly, traffic and iperf3 tests going over the VXLAN had much better throughput than traffic going directly over the Wireguard connection. IIRC, over the VXLAN, I was pulling ~800Mbps over the VXLAN/WG setup with iperf3.
Where this would fall apart is if there are firewalls in between that silently drop UDP fragments. In a case like that, it may be necessary to do VXLAN/Wireguard/Wireguard to conceal the fragmented packets with MTUs of 1500/1550/1440 respectively, assuming IPv4 and WAN MTU of 1500. I bet this would come with a significant performance hit though.
This is fun, but applications requiring L2 adjacency do it to limit latency/distance. Creating a L2 domain between here and the moon, what are you gonna use it for? Certainly not anything other than fun.
There's a number of specific scenarios this could be useful, like, some SANs can only replicate to L2 adjacent units. Say you wanted a replica off-site, and your gear is older/proprietary, you used to have to buy enterprise network gear to encap L2 and ship VLANs to remote sites. I wouldn't be dismissive of using VXLAN over wireguard to accomplish that.
I didn't say "this solution is easier", simply said someone might find a use case for it. Ya'll are dismissive of something neat with a number of use-cases (I only tossed out just one use case that popped into my head, based on actual experience on a million dollar SAN, that is still supported (and sold!) to this day).
If you have a million dollar SAN you have capital to get a real connection where you need it.
The reason I'm very against even discussing this is because people who don't understand the downsides would be open to doing this, shooting themselves in the foot along the way. People who understand the pitfalls just won't and are telling you DON'T, YOUR FOOT WILL HURT.
You could use VXLAN over WireGuard with a lower MTU, attach the VXLAN interfaces to different VRFs and route traffic, it's a somewhat valid usecase.
Switching over the internet is pain, I've got experience. Used to work at an MSP that did this as common procedure, worked fine until it didn't, and noone could explain why. And we're not even talking loops yet, you'll have to build a pretty sick RSTP.
The use-case is at best an SMB migration strategy.
You don't have to explain the perils of extending L2 over any type of WAN to me. There used to be a hard requirement from some SANs to have dark fiber for their replication - not just L2 adjacency (needed actual FC zones extended to another site if I recall). But all is not happy times with L3 links and BGP between things (and even then, BGP configs to achieve anything decent in terms of failover needs BFD or more). But sure, poop all over this fun thought experiment if you want, I doubt anyone's going to deploy this and put a billion dollar company at risk.
Yes you need the protocol designed to ensure bidirectional forwarding to keep failover times low. And the routing protocol for exchanging routes.
If the post mentioned the pitfalls it would be a fun experiment. But people don't have to ruin a billion dollar company to do damage, hopefully billion dollar companies can afford staff that knows this is wrong, not every SMB can.
Can you use this to get Apple bonjour / mDNS working over a remote network (connected via VPN)? Or similarly, could you use it for a cloud seedbox to cast to a chromecast on your local network (via the VPN obviously)?
For mDNS, you should run avahi to relay between subnets. For chromecast, that is SSDP/DLNA which is a multicast, so it is a matter of establishing mulitcast routing between sites.
Bridging L2 is not the optimal solution for either of your scenarios.
I took this to extremes last year: I used it to run MAAS from Australia to sweden (which requires layer 2). Granted I used tailscale to make the WireGuard part even easier, but it was a lot of fun.
Wow, I can’t believe the HN audience is so accepting of stretched layer 2 as a solution. It’s almost as though we’ve been invaded by middle management.
72 comments
[ 2.3 ms ] story [ 142 ms ] threadHow does it handle segmenting jumbo frames? Etc.
Edit: I see you changed your comment to consider remote DHCP which causes my comment above to be irrelevant. Oh well. The fact still holds that VXLAN is meant to handle Layer 2 so saying BGP is an alternative is like putting a square peg in a round hole.
For QoS, what do you think double-encapsulating all of the QoS bits inside of two tunnels is going to do for you? If you pay your carrier to respect QoS, you have to actually give them the bits! QoS goes on the OUTSIDE of the onion.
DHCP for VLANs should be handled by snooping for it on the switch and forwarding over L3 (ip helper-address), not by running vxlan over a VPN tunnel. And really it shouldn't ever be done over a potentially unreliable WAN link either; you should have a local agent for DDI at each site.
The only story here is that L2-over-L3 VPN have their very specific niches, and VXLAN+Wireguard is an alternative to L2TP+IPSec that is (IMO) both easier to configure and more reliable when such a thing is needed.
I'm sorry to ask but which alternative do you find easier and more reliable? I seem to be able to parse that either way, maybe it's just me.
Nope, here with the same problem and I don't treat L2TP/IPSec as an easier to configure and reliable.
L2 connectivity is still quite useful, even if you don't have a need for it.
I sometimes need to extend a system like this from one site to another. One is a calrec system (an audio mixer, I think it's the control traffic that needs to be sent), and I don't have enough access or time to see if I could build some kind of transparent proxy -- it won't work with multicast routing.
I do however have enough time to create a layer2 network between two nics. I tend to use mikrotiks for that, create an eoip tunnel (GRE with proprietary addons to cope with fragmentation) between the two endpoints and pop the interface in a bridge with a physical port, and move on.
"You don't need to do that thing."
How do you know?
Not sure if the author is reading this thread, but it's something you may find worth investigating fixing.
Mozilla/5.0 (Macintosh; Intel Mac OS X 10.15; rv:109.0) Gecko/20100101 Firefox/109.0
Code blocks without JS: https://i.imgur.com/Tcq34IK.png
Code blocks without JS under highlight: https://i.imgur.com/Lbvk2LS.png
Code blocks with JS enabled: https://i.imgur.com/kDCW4Q0.png
OpenBSD does support both routing domains and multiple routing tables and includes multiple routing daemons in the base system. I would recommend to the author to stop hacking at the keyboard, grab whatever not to structured visualisation tool works for them (e.g. a whiteboard, a block of paper, a random drawing app, Visio) and (re-)phrase the problem. Are you solving a problem or showing of how many acronyms you can expand without looking them up? This n layer encapsulation can work and can even be required to reproduce some (problematic) organisational structure, but it's far from elegant. Given the chance I would vastly prefer to just use multiple routing domains for the WireGuard tunnel interfaces and the underlay. It would result in far less complexity to manage as well as less overhead.
Why do so many people insist on tunneling Ethernet over IP? What's keeping operators from using IP routing (and just one layer of encapsulation) instead? Is IP routing so scary or everyone that indispensable applications that only work over Ethernet?
(also I think you lost one 'over UDP')
"wireshark can virtualize an L3 network out-of-the-box" How can this be true then?
This is essentially how you want to build your DC networks today too (why waste ips on netid and broadcast that you never use) (for public addresses).
But you stick BGP on there to exchange route information rather than static routing.
Each tenant can have multiple subnets; run BGP, etc. and the IP addresses can be re-used between different tenants.
It doesn't seem wireguard can do that with just p2p network. This was one argument I had for adding VXLAN encapsulation over IP for the above comment. You also get L2 connectivity if you want it.
As long as you're lowering MTU and not bridging your VXLAN you're fine.
You can run VXLAN over any L3 so it's versatile like that, but the benefit of VXLAN over other encap protocols is that it's simple, accelerated in a lot of hardware, integrated with BGP-EVPN to make scaling easier, more predictable and traceable.
I used to run a cluster of DNS/DHCP servers that were the first on site to run 5. The rest were 4.11 until we binned them for 6. Three cream coloured Compaq 3U lumps.
This one's running (present tense) on a 1 GHz Socket 370 Pentium 3. It's got some weirdness about Pentium 4 and newer CPUs. I think it can be patched up but the client doesn't want to pay for the work because "it's fine like it is." Not worth the headache to virtualize as the first attempt didn't work.
vMotion, Provisioning and backups work just fine.
Sometimes you just need an L2 tunnel. Most of the time you don't, but when you do, you do. For example, if you use IPv6 over SLAAC in a private network, you'll need to route NDP.
In the rare cases that you do need an L2 tunnel between two different locations, you probably want some kind of authorisation and authentication of the traffic to prevent injection/spoofing attacks and to make life just a but harder for the NSA (Google's use of HTTP was one way the NSA managed to tap connections that were otherwise encrypted by HTTPS). After all, this isn't just any traffic, these are internal subnets.
In terms of authorised traffic, Wireguard is quite lightweight and foolproof. Perhaps IPSec is even more lightweight but it's a pain to set up. The alternative would be to wrap all internal network traffic in an encrypted protocol and set up the necessary whitelists in the upstream ISPs.
The impact of such layering depends on the network connection between the data centers. If you can get jumbo packets across, fragmentation won't be a problem at all. If you run your own fiber between data centers, there's basically no downside until you're reaching very high saturation network saturation.
Where this would fall apart is if there are firewalls in between that silently drop UDP fragments. In a case like that, it may be necessary to do VXLAN/Wireguard/Wireguard to conceal the fragmented packets with MTUs of 1500/1550/1440 respectively, assuming IPv4 and WAN MTU of 1500. I bet this would come with a significant performance hit though.
The solution is getting a wavelength or dark fiber to the off-site, or throwing out the piece of junk SAN that only works on L2, it's too old by now.
The reason I'm very against even discussing this is because people who don't understand the downsides would be open to doing this, shooting themselves in the foot along the way. People who understand the pitfalls just won't and are telling you DON'T, YOUR FOOT WILL HURT.
You could use VXLAN over WireGuard with a lower MTU, attach the VXLAN interfaces to different VRFs and route traffic, it's a somewhat valid usecase.
Switching over the internet is pain, I've got experience. Used to work at an MSP that did this as common procedure, worked fine until it didn't, and noone could explain why. And we're not even talking loops yet, you'll have to build a pretty sick RSTP.
The use-case is at best an SMB migration strategy.
If the post mentioned the pitfalls it would be a fun experiment. But people don't have to ruin a billion dollar company to do damage, hopefully billion dollar companies can afford staff that knows this is wrong, not every SMB can.
Bridging L2 is not the optimal solution for either of your scenarios.
https://web.archive.org/web/20230214134248/https://rob-turne...
A little bit over my head, but an interesting read.
https://medium.com/@antongslismith/bare-metal-cloud-provisio...
Stretched layer 2 is almost always a mistake.