Thunderbolt Networking is FAST & CHEAP!

[@rafrackowiak]

This work of yours is spectacular!

[@Jims-Garage]

Thanks, glad it was useful! Anything you want to know give me a shout and I'll try to answer.

[@ob2522]

tmux is great for running stuff in the background, I use it all the time for switching between nodes in my cluster. Thanks for a fascinating vid!

[@Jims-Garage]

Thanks, tmux is on the list 👍

[@LaurenceHartje]

Thanks! I've been considering this exact same setup since the MS-01 was announced, but wanted to see some real life experiences before financially committing to this setup (as you mentioned thunderbolt Ethernet was in a strange state at that time). Looking forward to seeing the Ceph results and performance over the thunderbolt Ethernet.

[@Jims-Garage]

Thanks, glad it was useful. I'll be covering that part soon.

[@Mad-Jam]

Search on YT for a Video called "USB4 & Thunderbolt's TRUE Speed: Only 22Gbps!" He nailed it.

[@Jims-Garage]

Thanks, I will take a look at it.

[@scytob]

It's 26Gbps on intel NUCs, it varies on systems due to the DMA controller. The MS-101 only appears to be 20Gbps USB4 not 40Gbps TB4 which is why you see difference. It doesn't really matter as that 20% doesn't make much real world difference.

[@Jims-Garage]

@@scytob assuming that's because the 4 lanes are split 2x2 with ports? Traditionally you'd have a full 4 lanes?

[@scytob]

@@Jims-Garage i believe its a combination of how many retimers are used / Gen 3x1 vs Gen3x2 pcie. for example one could have a Genx2 but not implement enough retimers along with how much bandwidth is allocated to interdomain, vs tunneling and if it is fixed or dynamic on a specific implementation. Also shh USB3 is not USB (aka serial), it is actually a routed protocol that supports DP, USB3 and PCIE tunneling.... this is why your statement that thunderbolt networking was deprecated isn't correct, it just wasn't being well tested because they were focused on getting the core USB4 interdomain routing working first, i found some critical interdomain bugs that needed to be fixed.... and they fixed the IPv6 bugs as part of that, anyhoo, this is all based on my (mis?)reading of the USB4 specification documents. Bandwidth allocation is documented on page 60 of the connection manager guide for USB4 - its a heavy read, lol, also look at the interdomain service guide to see why thunderbolt networking was absolutely foundational too.

[@scytob]

@@Jims-Garage or to say it another way the bandwidth allocation between pcie / usb3 / dp tunneling etc is done dynamically across all ports - the lanes are not split per port - its the total number of lanes that matter, nothing else (unless someone doesn't something silly like one thunderbolt controller per port....)

[@squidiebah]

Thanks for the video, I've been wanting to use USB4 between two Linux cluster nodes for a while and was curious about stability. Thunderbolt-net module supports networking thankfully.

[@Jims-Garage]

Great to hear! Let me know how that goes. It's a great way to have fast, "cheap" networking.

[@uvalleza]

Jim, in your running config i am not seeing the "ip router openfabric 1" only the "ipv6 router openfabric 1" one, which explains why you cant use ipv4 (this is in the interface en05/en06 section when you showed "show running-config"). But fyi, i went through this like 3 times because for some reason it sometimes doesn't take the setting, so have to redo it until it picks it up. I did this for my UM790 Pros and have MS01's on the way! So cant wait to set it like this :) but just discovered your channel and loving the content man. Keep it up!

[@Khaldrogo5]

Happened to me too. first time typing "ip router openfabric 1" for the FRR config gives an error "Couldn't bring up interface, please check log.". If you just re-enter "ip router openfabric 1" the error goes away and get added into the config.

[@scytob]

nice find, doing cat /etc/frr/frr.conf will show what frr has comitted (note NEVER edit this file if it is wrong, use the frr console)

[@scytob]

i made some edits to the gist that might avoid this, its unclear why this broke on some systems - i wonder if it is a bug - the key is use vtysh -c "show rrunning-config" will show if the confi took or not

[@scotteecarr]

This is an incredible bit of kit. I wish these MS01s existed when i built my cluster. I'm pretty sure you'll need more then 3 nodes if you plan to run ceph as 3 is the bare minimum count. You don't want split brain, and 3 nodes won't allow for you to perform maintenance.

[@scytob]

works fine with one node down - i found that out the hard way, lol, stayed working, performance was great and gave me time to get node 3 back up and running

[@LtdJorge]

There is no split brain with 3 nodes. As long as you have it set up as size = 3, min_size = 2, it will run with 2 nodes up in degraded mode. If there was a network split where you’d get 2/1 nodes separated, the group with the 2 nodes can continue running while the group with 1 cannot.

[@BZFFirst]

Hi. Would you do a video on how to configure the SR-IOV for the Miniforums. And how to use it in VM's?

[@Jims-Garage]

Yes, I'll likely cover in the near future

[@scytob]

@@Jims-Garage even i want this, i cant get that to be reliable.....

[@simo47768]

Wow. Amazing video again.

[@Jims-Garage]

Glad you enjoyed it!

[@jacobburgin826]

Most of us are struggling with IPV4 Especially after reboot. I ended up taking it out and just using IPV6 Restarting the fee service usually brings ipv4 back up but rebooting the system kills it again

[@scytob]

IPv4 takes longer to converge, i never figured out why as IPv6 is better suited to this task IMO

[@antonio.luevano]

How did you get the lo:0 and lo:6 to show up in the GUI? I followed the instuctions and the thunderbolt is working, but is not being seen in the UI. Great tutorial. TIA.

[@Jims-Garage]

@@antonio.luevano I will have to check. Since recording there has been a Proxmox update and changes to the methodology.

[@antonio.luevano]

@@Jims-Garage I appreciate your response. I really enjoy your videos, you are very through with the steps and concepts.

[@ryanmalone2681]

Like the sweater, but I’m really missing the pastel cardigans. Video isn’t bad either 😉. Appreciate you.

[@Jims-Garage]

Haha, thanks. I will try to sort out my attire ASAP

[@francoismartin5578]

Hello Really Thank’s for this great video ❤

[@Jims-Garage]

Thanks for visiting

[@mihaitamas]

For max performance you will need to set affinity on the performance cores for iperf3. That will get you to 26Gbits/s across the board.

[@Jims-Garage]

I believe for the MS-01 it's due to the PCIe lanes being shared between ports and DMA, not CPU affinity.

[@mihaitamas]

@@Jims-Garage I highly doubt it as before setting the affinity could not get more than 18-21Gbps and with a LOT of retries on one of the nodes, like >8K retries on a 60 sec iperf test. What I am trying to say, is that you might be lucky, but better to chose and have consistent performance, rather than try to guess what's wrong. ;)

[@bartomiejlesniowski8635]

@@Jims-Garage , @mihaitamas's right, try use this on each node: #!/bin/bash for id in $(grep 'thunderbolt' /proc/interrupts | awk '{print $1}' | cut -d ':' -f1); do echo 0f > /proc/irq/$id/smp_affinity done

[@lmaguire]

Pretty sure your V4 issue is the fact that you’re trying to set a/32 on those if you go back to old school sub netting and remember how Anding works. I believe it’ll pop up that each of those is going to think it’s on its own isolated network.

[@scotteecarr]

I agree. Change the /32 netmask to /24 or something to include more addresses.

[@scytob]

the advantage of a /32 is it fixes the subnet as a single IP - its a very effective strategy to hard code an IP when one can't.

[@akurenda1985]

That's some great network speed, but I can't help but see you're just using local LVM storage. So no zfs replication or ceph? Seems like a waste to just use the 25 gigabit network just for migrating VM's.

[@Jims-Garage]

Thanks. As I mention in the video I'll be doing Ceph in the near future (want the infrastructure in place first).

[@ajayganeshie1857]

Hi there Jim, how are you? Is this also possible with fast usb? Thank you! Ajay

[@Jims-Garage]

@@ajayganeshie1857 I believe USB 4 as it's thunderbolt 4 (otherwise I haven't tested as I don't have the means to).

[@潜水屋]

is Thunderbolt Networking working between mac and pve, or between windows and pve.

[@Jims-Garage]

Good question, I don't know. I suspect it could do but I'm only using it for networking between nodes and I don't have Mac to test with

[@NerdzNZ]

I am looking to do this with my MS-01s but I am not running them with Proxmox, I am using Debian 12. Any chance you have notes on what would be different for that approach?

[@Jims-Garage]

@@NerdzNZ the process should be identical, Proxmox 8 is Debian 12 under the hood.

[@NerdzNZ]

@@Jims-Garage This bit tripped me up, because those are very proxmox NIC names: allow-hotplug en05 iface en05 inet manual mtu 65520 iface en05 inet6 manual mtu 65520 allow-hotplug en06 iface en06 inet manual mtu 65520 iface en06 inet6 manual mtu 65520 Mine in the MS-01 looks more like enp87s0 so I wasn't sure if it was "safe" this is new territory for me and network is not a strong point

[@Jims-Garage]

@@NerdzNZ eno and enp is just the naming convention depending on how the devices are detected. Just amend the instructions accordingly. Should be fine.

[@MarkConstable]

I've been using Ceph on my little mongrel cluster since early last year and it's been fine on a 2.5GbE fabric (started off with 1GbE). BTW, you mentioned passing through a GPU to containers in a k3 cluster so I suspect you are using SR-IOV and not full iGPU passthrough.

[@Jims-Garage]

For now it's full passthrough with 3 agent nodes. I will be looking into SR-IOV but with kernel 6.8 the custom DKMS doesn't work.

[@MarkConstable]

@@Jims-Garage Cool. It would be really exellent if you could detail your exact settings somewhere because I'm not aware of anyone else that has cracked this particular procedure.

[@cberthe067]

Does SR-IOV is able to share a GPU among VM/LXC container ?

[@cryptodendrum]

I wonder how Apple Thunderbolt4 cables would perform? If that would cure your retransmission error count? Or if the observed netlink speed would be any faster?

[@Jims-Garage]

I can't imagine it would make a difference, it's a reputable cable that meets the specifications. The speed is likely limited by 2x PCIe lanes and DMA.

[@cryptodendrum]

@@Jims-Garage Thanks for the advice. My Cable-Matters branded Thunderbolt cables arrived, along with my RAM and NVMe's. Just waiting for the computers for our new cluster to arrive now. Looking forward to getting started with it. FYI - it'll replace an old Protectli i5 unit from 2018 & two old MacMini's; all which run XCP-NG currently and I plan to try to run XCP-NG on my new MS-01 cluster.

[@JonatanCastro]

I actually tried with my Apple Display studio cable and it's same speed and retries

[@Jims-Garage]

@@JonatanCastro good to know

[@NetBandit70]

My-craw-tic Wat!?!

[@Jims-Garage]

How is it supposed to be pronounced?

[@xenos1983]

Hi Jim Thank you for your very comprehensive video. I really appreciate your work. Do you know if this will also work on Thunderbolt 3 on older NUC7i5BNB?

[@xenos1983]

After some testing I've gotten it working. Had trouble to obtain the correct pci path because it was nested in a longer path on my old NUC (12Gbps). This command gave me the correct result for the path: udevadm info /sys/class/net/thunderbolt0 | grep ID_PATH

[@shephusted2714]

do some benchmarks

[@Jims-Garage]

What would you like to see?

[@TheMongolPrime]

Love the node names. I knew I liked you.

[@Jims-Garage]

Haha, thanks! For The Emperor and Sanguinius!

[@michaeljolley6773]

This may be a dumb question but can I do this using USB 3.1 Gen 2? I don't have USB C ports but at least the USB 3.1 Gen 2 has faster speeds than my single 1gb lan port on each node.

[@Jims-Garage]

I don't believe so, you can do it with thunderbolt 3 though I think.

[@scytob]

USB3 doesn't support XDOMAIN specification so cannot use XDOMAIN networking (aka thunderbolt). USB4 is actually thunderbolt 3+some extensions (TB4 has more mandatory items).

[@scytob]

@@Jims-Garage @scytob 0 seconds ago USB3 doesn't support XDOMAIN specification so cannot use XDOMAIN networking (aka thunderbolt). USB4 is actually thunderbolt 3+some extensions (TB4 has more mandatory items).

[@crc-error-7968]

Ciao Jim, which is the power consumption of the 3 ms01? idle, with few wms, etc.?

[@Jims-Garage]

This cluster is currently pulling 160W with what you see. I suspect you could get lower with core pinning but I haven't done that yet. I simply wanted to migrate over ASAP as that would give me power savings and speed improvements regardless.

[@crc-error-7968]

@@Jims-Garage So, about 54w/unit. Honestly, I thought they would consume much less when idle.. I'm looking for a solution similar to yours to downgrade (in terms of consumption) and upgrade (in terms of CPU) my homelab, but it's easy to make a mistake since there is an ocean of possibilities between mini PCs and components to choose from.

[@Jims-Garage]

Idle consumption is almost pointless because if it's idle turn it off. The figures I gave were running around 12VMs across all nodes.

[@crc-error-7968]

@@Jims-Garage sorry, for idle I mean do nothing heavy, something like mqtt receiving data, home assistant does its stuffs, uptime kuma which monitoring services, traefik waiting new requests to redirect, jellyfin waiting me choosing something to see, etc.. etc..

[@ClacKeyTech]

i recently built 2 pro aggregation switches in my schools network

[@Jims-Garage]

That's awesome. Must be great for that kind of environment.

[@ClacKeyTech]

@@Jims-Garage absolutely

[@cberthe067]

When setting Ceph, use erasure-coding and not replication to maximize disk space ... may possible that you need to create it manually with shell command as if i remember Proxmox only support replication ...

[@scytob]

erasure-encoding has its downsides in terms of latency and reliability in a 3 node cluster

[@ws_stelzi79]

Man how you pronounce Mikrotik ... 🤔

[@Jims-Garage]

How's it supposed to be pronounced? 😂

[@headlibrarian1996]

Micro-tick was my assumption.

[@cryptodendrum]

@@headlibrarian1996 I've always heard it pronounced Micro-tik too, but from now on, it shall officially always be called My-craw-tic. I'm going to use that everywhere now. :D

[@JeramiFrost]

This may be inconsequential, but at roughly 25:50 you are talking about using the IPV4 link you created, but the actual page that shows your migration settings shows the IPV6 address selected. Does this matter? Edit: After going back to the beginning of the video and doing the ping test i answered my own question. I could not ping with the IPV4 selection. I switched it to the IPV6 like yours showed and i was able to successfully ping. Thanks for an awesome, and super helpful video.

[@Jims-Garage]

You're welcome, thanks 👍

[@Benjamin-rd7xi]

Hey there, I noticed something at kzbin.info/www/bejne/ipOWfHtoh8emiqc where you show the topology. I the guide it shows that the interface for the other nodes switch between en05 and en06. In your topology you are only using en06 this the next hop to sanguinius is dorn, but it really should be sangunius. So if dorn is down your connection betwenn the other two will be broken. Maybe this also helps with IPv4

[@Jims-Garage]

@@Benjamin-rd7xi thank you, I will double check that to see if it's still an issue

[@djmillhaus]

You're using TB4 cables, but I wonder would USB4 40Gbps cables that are "compatible to TB3" do that job as well?

[@Jims-Garage]

AFAIK tb4 and usb4 are the same.

[@simonthornroos752]

@@Jims-Garage @djmillhaus I actually tried with usb4 cable and those didn't get recognized by the ms01 on any of my 3 nodes. I have just placed an order on tb4 cables. Fingers crossed those will work!

[@Jims-Garage]

@@simonthornroos752 good to know, thanks

Thank you for breaking it down. Answered all the questions I didn't even know I had!

[@Jims-Garage]

Glad it was helpful!

[@TradersTradingEdge]

Holy smoke :) That's awesome. Thanks for that.

[@Jims-Garage]

Thanks 👍

[@fasti8993]

Can you provide information on cpu utilization if the Thunderbolts are doing maximum traffic? If I get it right, you are kind of sacrificing a little bit of your cpu as an ethernet controler… This might play a roll for making a decision on which version of the machine to by cause dual a 25g nic goes for 50 bucks on ebay…

[@Jims-Garage]

I'll take a look at it and post online

[@djvincon]

FOR THE EMPEROR!

[@Andy-fd5fg]

hummm mesh networking... now all i need is enough computers to try this on

[@Jims-Garage]

I was really surprised with how good it is!

[@jgarfield]

What do you use for diagramming?

[@Jims-Garage]

This is draw.io

[@BenjaminBenStein]

🎉

[@totoro1596]

cool stuff! why lvm instead of zfs?

[@Jims-Garage]

For speed, this is the only 4x4 slot. I also backup daily and replicate so infrastructure isn't a problem.

[@ewenchan1239]

Thunderbolt networking is cheap if you have three MS-01s, which are at least $439 USD barebones or $629 USD for the Core i5 12600H version with 32 GB of RAM and a 1 TB SSD. But if you want the top of the line, pre-configured model from Minisforum, each node will set you back $829 USD, which means that three nodes + Thunderbolt cables, will run you closer to $2500 USD in total.

[@Jims-Garage]

True for the MS-01 but many consumer devices have a couple of thunderbolt ports. It also means you don't have to buy additional adapters.

[@ewenchan1239]

@@Jims-Garage Varies. Depends on what it is, and its respective age. My 7th gen NUC - no. 8th gen NUC has one Thunderbolt 3 port on it. None of the rest of my systems, not even my Asus Z690 Prime P motherboard (for 12th gen Intel) has Thunderbolt on it. So it REALLY depends.

[@scytob]

For folks who already have decided to use NUC with TB form factors this is basically 'free' compared to buying 3 TB 10gbe adapters....

[@scytob]

@@ewenchan1239 and none of that invalidates its basically for free for folks who are making their purchase decisions now...

[@ewenchan1239]

@@scytob If you have a NUC that only has a single TB port, then you won't be able to do this (per Wendell) in terms of being able to set up a near token ring type of network using nCr(3,2) combinations of ports on systems. The best that you'd be able to do with a single port on each NUC is a single pair of point-to-point connection, which means that the 3rd system will NOT be connected to said Thunderbolt-based high(er) speed network.

[@casperghst42]

As for the drop outs, Wayne Fox made some testes using Caldigit and Apple (Pro) cables, and found that non-active cables might not always be perfect (kzbin.info/www/bejne/eJS7aaybpsuSa9ksi=siTWqYroIBC6fzL4) - there is a price difference though.

[@scytob]

I use short OWC TB4 cables, work perfectly.

[@nirv]

Why is there a dollar sign in the center of a heart and "thanks?" Why are there affiliate links in the description? No dude. You are violating the spirit of the internet. Get these dollar signs out of my face. This is the INTERNET. What are you doing dude?