• 0 Posts
  • 29 Comments
Joined 3 years ago
cake
Cake day: January 16th, 2024

help-circle


  • Seriously I couldn’t even get my wife to use a third party messaging app.

    Thank God iPhones finally support…RCS.

    The problem is, if we fold on our kids, we are essentially just handing them over to meta. Forever. They will end up being as locked in and helpless as every other adult.

    Facebook is the new smoking. Even worse. Because parents are now expected to put the cigarettes into their kids mouths themselves…and if they don’t? Your kid will be a social leper, and have the most miserable adolescence available.

    Parents can’t even feign ignorance. We know how bad social media is…but we’re addicted. They’ve got dopamine on a treble hook right into our brains and they know it.

    That’s not even teenage hyperbole. Like that’s legit what life is like for kids. No socials? May as well not exist.


  • Lenovo T-series is generally a good pick whenever you want to run Linux.

    I’d say the same for their USFF and SFF form factors as well. Great systems for homelabs, I have several (and my laptop is a T495, and my son has my old T470s).

    Problem with laptops and (u)SFFs is storage capability. If you want a NAS you need some sort of external JBOD or you will be limited to 1-2 storage disks (and possibly have to boot I.e. truenas from USB).

    If you can find them with a PCIe slot (or extra m.2 slot that exposes 4 lanes at least), you should be able to get a RAID card and rig up a way to use eSATA and some sort of external disk enclosure though.

    Pay attention to performance/watt. Consider different boxes for different roles if you want…I.e. ZFS doesn’t need a lot of processing power, but it does like RAM. If you have a system that can use dirt-cheap DDR3(L) like a 6th or 7th gen Intel, that may be a better choice for NAS-only.

    Then use a more powerful NUC or (u)sff for proxmox and run LXCs or VMs on it.



  • JasonDJ@lemmy.ziptoSelfhosted@lemmy.world•Is anyone self-hosting their own AI?
    link
    fedilink
    English
    arrow-up
    26
    arrow-down
    2
    ·
    1 month ago

    What a timely comment.

    I recently gave my kid a gaming PC for his birthday…which I used as an excuse to upgrade my GPU from a 1660 Super to a 5060ti 16GB

    Combined with a 5600x3d and 32gb of system memory.

    I just started messing with local LLMs yesterday. “Getting started” is rather easy. pacman -S lm-studio (I use arch, btw), launch it, download a model that your hardware supports, and start talking.

    Right now I’m working on getting it to talk to vscode.


  • Router is a bunch of old PCs I frankensteined and put opnsense on.

    Ended up i5-7500, 8gb mem, 500gb SSD.

    Put in a dual-port SFP+ NIC and a dual port 2.5GbaseT nic.

    One of the 2.5gb ports goes to my cable modem.

    One of the 10gb ports takes Twinax to my dumb LAN switch, the other to my dumb server switch, which right now is really just my storage network…the link to the firewall isn’t doing anytuhing.

    The dumb switches I have here, though, they’ve got some sort of weird MAC learning disability. It cannot handle macflaps, which I found out about because I had a hard time roaming between APs.

    So I got a Netgear GS108T and put OpenWRT on it, and that switches between my APs.

    My APs are Linksys MX4200v1’s. Also openwrt. Got four of them…one upstairs, two in the ground floor, one in the basement. Can get very close to 1Gbps anywhere in the house on wifi. Not the best platform for openwrt (ipq807x), as I’ve found out…but it was a hella cheap deal for tri-band AX APs.



  • It’s okay for small scale, dev, etc…on a fast and flat/local network.

    But databases really don’t like NFS, SMB, or most file sharing protocols. There’s too much transaction overhead.

    The effect gets exponentially worse as you add in latency. Database may as well be on the moon if you’ve got more than 10ms between it and the app (hyperbole but not really).

    Use fast local storage, ideally. Or block if you have it. Async replicas if you are chasing 9s.

    When databases fail, they fail spectacularly and cause ripples. Remember that you’ll be the one supporting it when they do.



  • This is one of the right answers, and probably the easiest.

    But…who is your domain registered with? A lot of the big registrars (I.e. CF) can easily integrate into several opnsense plugins…namely ACME/HAProxy, and Dynamic DNS.

    Dynamic DNS would make sure that “pubip.mydomain.com”, from the outside, always pointed to your Internet connection, even if your IP changes.

    Then just register CNAMEs for your external services to point to your pubip.mydomain.com.

    HAProxy reads the CNAME and directs it to the right service.

    You can either use split DNS and have HAProxy listen on both the internal and external, and map DNS around that way…or you can just listen on your external interface and connect that way and only update one end.

    Imo this is the more comprehensive solution, because this also easily solves certificates (via ACME). The last leg (from HAProxy to backend) and be unencrypted or just use long-life self-signed certs if you really care…but I wouldn’t worry about encryptipn on the inside of a homelab.



  • You are probably fortunate enough to live in an area with not a lot of wireless interference, or don’t have a lot of wireless devices on your network, or all your devices are rather modern.

    You can have good wireless…it won’t be as good as wired, never will or can be. By it’s very nature it has to be a shared medium, which by it’s very nature must be half duplex. Then there is the matter of sharing time with other stations on the same channel (and not necessarily on your network or under your control).

    But for most things it can easily be “good enough”. Certainly wouldn’t run a server on wireless. Generally worse performance for realtime voice/video/gaming as well, because latency and jitter matter, and by having to share the channel with other stations, jitter is a fact of life.

    And unsteady wireless is the #1 cause of VPN complaints in my office. By a very wide margin. Miraculously the problem always goes away as soon as they plug in.

    But for casual web browsing, occasional large downloads (there’s a reason mobile apps don’t really show you how many B/s your transferring), streaming media…yeah, wireless can easily be good enough.


  • Yeah the SFFs and uSFFs are great for consuming low power.

    I’ve got a couple m910q’s that I bought for htpc purposes but honestly I might be better off putting the thick ones on htpc duty. Just gotta work out some kinks with how they handle suspend when hooked up to a TV.

    They’ve really shot up in price…I bought two for 130 shipped (total) with i5-7500T and 16gb back around last Halloween. I’m seeing singles not specced as well for the same price

    The tough thing about uSFF is that they usually don’t have any PCIe slots…so if you need a second NIC, or want to upgrade to mGig, you can’t, or they have to be over USB 3.


  • This.

    I look for corporate desktops that are off-lease or EoL. Big fan of Lenovo M series. 6th/7th gen still have lots of life left in them and plenty of power for most homelab tasks.

    Anything much newer than that will consume less power per core (usually) but will cost more up front and probably be more expensive to put upgrades into (i.e. DDR5). Anything much older than that won’t be worth the performance per watt. By the time you put a 7th gen through its paces, you’ll be ready to upgrade and have a much clearer look at what you want/need.

    Right now I’ve got 3x M710s forming a kubernetes cluster, and another running opnsense only.

    Kubernetes was built on top of VMs in proxmox, but I’m thinking I will move them to metal.

    I also have what was my PC down in the basement, with proxmox, running a TrueNAS VM and a Bazzite VM for GPU passthrough and Sunshine, but I’m gonna reclaim those guts, put a spare 6th gen with DDR3 and run TrueNAS on metal. Then I’ll have my PC back.




  • Netbox is a documentation tool. You can plug in Napalm to do some stuff but it mostly exists to catalog the intended state of the network.

    It’s a wonderfully powerful tool, and Stretch has done a great job with it…but it’s not an analysis tool, it’s documentation.

    Stretch is a pretty cool guy too. He strikes me as the kind of person that really wants to help colleagues “see the light” of the role Python and FOSS can play in network automation and maintenance. I respect that, a lot…finding enjoyment in the way you do things, and wanting to share that with other people.



  • Yeah I’d check for fragmentation, particularly coming from whatever was on the opposite end of this tunnel. This looks like librespeed (which is super simple to run in a container, ‘adolfintel/speedtest‘, if interested…I run some at work and it’s very useful) so I’m assuming it was running on the server at the other end of the wireguard tunnel?

    That latency and jitter are also absurd tho. Op should run a bufferbloat test on both sides. Though I don’t always trust those results from librespeed.