Right? Isn’t that post they way most of us did it but we also did it worse?
Anyone else get into the habit of running a second sshd or dropbear after the second time they took out ssh on something that didn’t have a remote console? Fun times.
Seems like a wonderful start for colocating stuff. Sadly “throw a bunch of money” normally means “throw a bucket of money” - which is too much for most individuals.
I think about colocating some small machine at a friends company sometimes and I remember that I have gigabit at home at their office has almost no internet service at all.
When I was a PhD student, I used a company called Mac Mini Colo. They were based in Texas (I was in the UK) but offered remote hands for install and emergency use, and I managed to both install OpenBSD and recover from a botched upgrade with them. They were taking advantage of the fact that the newly launched Mac Minis were tiny in comparison to a 1U server and so built custom racks to host them. It wasn’t server-grade hardware (though it was fun having a PowerPC machine in a colo when everyone else was using x86 VMs).
Blade servers were meant to revolutionise this kind of thing by providing very small rack-mounted things. I’ve seen them in racks but I’ve rarely seen a place offering to sell you hosting by the blade. I think VPSs ate most of the cost-sensitive market. The cheapest vultr VM is $2.50/month ($3.50/month if you want IPv4 as well as v6). You can’t really get close to that with dedicated hardware. For a while, there was a ‘physicalisation’ buzzword. I remember one vendor providing a package-on-package (PoP) stack that was a Cortex-A8 CPU, 1 GiB of RAM, and some amount of flash that I don’t remember. There idea was that you’d put these on a big board with access to a shared network (and possibly iSCSI for remote storage if the flash wasn’t enough) and be able to rent them for a few dollars a month as dedicated machines. The problem was that, although they scaled down well, they didn’t scale up: you couldn’t easily combine them into larger machines, whereas buying a 32 core x86 machine gave you the ability to sell anything from oversubscribed 1 VCPU VMs up to uncontended 32 VCPU VMs. Similarly, if you have a big iSCSI array that you expose as virtual disks, you can easily offer anything from 1 GiB to 100 TiB (or larger) slices.
OVHCloud’s Kimsufi range seems to offer some very cheap dedicated servers, but you don’t get to provide the hardware (so it’s not a true colo). They’re recycled from more expensive tiers. That said, the cheapest ones start at $5/month, so if they’re okay for you then they might be a good alternative. I’d be a bit nervous about the disks, given the age of the hardware though.
My website is hosted on a friend’s server which is currently located at Mythic Beasts in Cambridge. Cost is the same ballpark as a dedicated server from a cheap & cheerful hosting provider. But I would avoid hardware unless there’s a really good reason: for most purposes a VM will do the job, and it will be cheaper and more resilient. (Recently my own VM had to move to a different building, which required no effort from me nor any noticeable outage.)
What I discovered while working in the colo space is that many companies deprecate their hardware at quite a fast rate, oh my super powerful blade system is 5 yo, let’s buy a new one and … well, what’s are we doing with that? And they would just sell it cheap.
If you have access to that kind of market somehow (sometimes it even reaches eBay), then you get great stuff for cheap. Be careful as sometimes you get very bad deals (oh the actual port 20-24 of the switch never worked, forgot to mention that…).
I kind of like the model where somebody else manages the hardware and just gives me access to the bare metal. That way I don’t have to flip out anywhere in the middle of the night when my stuff blows up. I don’t mind reinstalling an OS, but I really hate racking servers.
The main thing I like about VMs relative to this is that someone else owns disk failures. Most cloud providers these days don’t oversubscribe VCPUs. I think in Azure the B class does but everything else is on CPU per VCPU (B class is then padded in around the edges and runs at lower priority, so anything else can use 100% of the CPUs if it wants to, it just turns out that most customers don’t and so there’s space for the cheap ones). Almost no one oversubscribed memory (except for container workloads where disk cache for common container images can save a lot of memory). So the benefit of a dedicated server is small unless you’re doing something very latency / jitter sensitive. In contrast, most cloud providers have built a reliable storage back end (lots of error correction and redundancy) that they use to implement VM disks. It is very likely that at least one disk will die in a storage pool that’s backing your VMs, but you will never be aware of it (unless there’s something that causes catastrophic failure of a large chunk of a storage rack, and even then you won’t if you’ve paid for the higher levels of replication). Disks failing in a dedicated server are your problem and so you’re responsible for RAID and so on, and you need more redundancy than if the reliability can be amortised over more disks and you have to deal with remote hands to replace a disk if it fails (I’ve had a disk fail in a colocated server and it was annoying).
I pay about $50 a month for a server with 24 course and 32 GB of ram. The nice part is the hard drives which are half SSD which are $120 gigs and a two terabyte 2 terabyte magnetic disc. It’s older hardware but I get five IP addresses and a/64 IPv6
I feel old.
Right? Isn’t that post they way most of us did it but we also did it worse?
Anyone else get into the habit of running a second sshd or dropbear after the second time they took out ssh on something that didn’t have a remote console? Fun times.
Followup to this post: https://lobste.rs/s/j85ynw/terrible_way_jump_into_colocating_your
Seems like a wonderful start for colocating stuff. Sadly “throw a bunch of money” normally means “throw a bucket of money” - which is too much for most individuals.
I think about colocating some small machine at a friends company sometimes and I remember that I have gigabit at home at their office has almost no internet service at all.
When I was a PhD student, I used a company called Mac Mini Colo. They were based in Texas (I was in the UK) but offered remote hands for install and emergency use, and I managed to both install OpenBSD and recover from a botched upgrade with them. They were taking advantage of the fact that the newly launched Mac Minis were tiny in comparison to a 1U server and so built custom racks to host them. It wasn’t server-grade hardware (though it was fun having a PowerPC machine in a colo when everyone else was using x86 VMs).
Blade servers were meant to revolutionise this kind of thing by providing very small rack-mounted things. I’ve seen them in racks but I’ve rarely seen a place offering to sell you hosting by the blade. I think VPSs ate most of the cost-sensitive market. The cheapest vultr VM is $2.50/month ($3.50/month if you want IPv4 as well as v6). You can’t really get close to that with dedicated hardware. For a while, there was a ‘physicalisation’ buzzword. I remember one vendor providing a package-on-package (PoP) stack that was a Cortex-A8 CPU, 1 GiB of RAM, and some amount of flash that I don’t remember. There idea was that you’d put these on a big board with access to a shared network (and possibly iSCSI for remote storage if the flash wasn’t enough) and be able to rent them for a few dollars a month as dedicated machines. The problem was that, although they scaled down well, they didn’t scale up: you couldn’t easily combine them into larger machines, whereas buying a 32 core x86 machine gave you the ability to sell anything from oversubscribed 1 VCPU VMs up to uncontended 32 VCPU VMs. Similarly, if you have a big iSCSI array that you expose as virtual disks, you can easily offer anything from 1 GiB to 100 TiB (or larger) slices.
OVHCloud’s Kimsufi range seems to offer some very cheap dedicated servers, but you don’t get to provide the hardware (so it’s not a true colo). They’re recycled from more expensive tiers. That said, the cheapest ones start at $5/month, so if they’re okay for you then they might be a good alternative. I’d be a bit nervous about the disks, given the age of the hardware though.
[Comment removed by author]
My website is hosted on a friend’s server which is currently located at Mythic Beasts in Cambridge. Cost is the same ballpark as a dedicated server from a cheap & cheerful hosting provider. But I would avoid hardware unless there’s a really good reason: for most purposes a VM will do the job, and it will be cheaper and more resilient. (Recently my own VM had to move to a different building, which required no effort from me nor any noticeable outage.)
What I discovered while working in the colo space is that many companies deprecate their hardware at quite a fast rate, oh my super powerful blade system is 5 yo, let’s buy a new one and … well, what’s are we doing with that? And they would just sell it cheap. If you have access to that kind of market somehow (sometimes it even reaches eBay), then you get great stuff for cheap. Be careful as sometimes you get very bad deals (oh the actual port 20-24 of the switch never worked, forgot to mention that…).
I kind of like the model where somebody else manages the hardware and just gives me access to the bare metal. That way I don’t have to flip out anywhere in the middle of the night when my stuff blows up. I don’t mind reinstalling an OS, but I really hate racking servers.
The main thing I like about VMs relative to this is that someone else owns disk failures. Most cloud providers these days don’t oversubscribe VCPUs. I think in Azure the B class does but everything else is on CPU per VCPU (B class is then padded in around the edges and runs at lower priority, so anything else can use 100% of the CPUs if it wants to, it just turns out that most customers don’t and so there’s space for the cheap ones). Almost no one oversubscribed memory (except for container workloads where disk cache for common container images can save a lot of memory). So the benefit of a dedicated server is small unless you’re doing something very latency / jitter sensitive. In contrast, most cloud providers have built a reliable storage back end (lots of error correction and redundancy) that they use to implement VM disks. It is very likely that at least one disk will die in a storage pool that’s backing your VMs, but you will never be aware of it (unless there’s something that causes catastrophic failure of a large chunk of a storage rack, and even then you won’t if you’ve paid for the higher levels of replication). Disks failing in a dedicated server are your problem and so you’re responsible for RAID and so on, and you need more redundancy than if the reliability can be amortised over more disks and you have to deal with remote hands to replace a disk if it fails (I’ve had a disk fail in a colocated server and it was annoying).
What are the prices for this these days? Is it possible to get a reasonable overview of the space?
Depends what “this” is, but colo space is cheap/expensive depending on how much you are using. E.g.: https://www.hetzner.com/colocation/
In German I just read this which has a bunch of truth but still no apples to apples comparisons: https://blog.koehntopp.info/2024/09/30/cloud-cost-vs-on-premises-cost.html
I pay about $50 a month for a server with 24 course and 32 GB of ram. The nice part is the hard drives which are half SSD which are $120 gigs and a two terabyte 2 terabyte magnetic disc. It’s older hardware but I get five IP addresses and a/64 IPv6