I rent a small Linode to run a minecraft server for my extended family. Sometimes it struggles to keep up with just a half dozen of us, and when I tried swapping from a spigot based system to a forge one so the kids could have mods it was basically unplayable. I can't imagine scaling out to thousands of users, this was some amazing work.
While developing this did you come up with any best practice recommendations for individual nodes?
Your single-core performance is the most important thing for a Minecraft server, especially for Forge mods.
For vanilla, I’d try Paper (a high performance spigot fork) and see if you still have problems. If you’re lagging with 6 people while running Paper, you simply need a better CPU.
Tangentially related, I had to move a web app off of DigitalOcean because the CPU performance was terrible compared to what you get for $80/mo from a dedicated server from OVH. I know Linode and DigitalOcean compete. Get a OVH dedicated server, try the CPU passmark bench or whatever and compare. AWS + Linode + DigitalOcean are all VPS and pretty slow.
It might be too pricy, but if you go to about 25usd/month, you can get a dedicated server with an i7 from hetzner server auction. These cheap boxes don't have ecc ram etc - but should work well for things like this.
On the page, one option advertised is for €23.53/month with an Intel Core i7-4770, 2 TB HDD, and 32 GB RAM. What’s the catch? How is it possible to rent such hardware for so cheap?
There is actually a catch (besides very little customer service, which is par for the course with cloud stuff anyway), but it's not really their fault. Hetzner doesn't peer with a lot of bigger players, and therefore its datacenters can have really lousy / inconsistent bandwidth and latency for people outside a relatively small area in Europe (and can occasionally have bizarre outages that are presumably related to partitions). If that's not a dealbreaker for you, then they're one of the best deals out there if you don't want to manage your own hardware and don't need really efficient access to other cloud services like S3 (which obviously work much better within Amazon's datacenters).
Ehh you can get cheaper compute in Fremont if you're willing to rent >=1 rack. I'm guessing you only think €23.53/mo is cheap because you've been getting reamed by AWS pricing :P
I'm not sure what you imply with the quotes around dedicated servers - this is hardware for rent, not VMs (hetzner has VMs too, starting from about 5usd/month - but that's something else).
You do need to go a bit up in price for proper server hw - although there are a few aging xeon boxes with ecc ram on the low end now.
They are second hand servers - if you want brand new servers you pay a setup fee, and more pr month. Once you upgrade, or cancel, that (once new) server goes into the auction at a lower price.
If you go with Oracle Cloud, you can get an 4-core, 24gb RAM aarch64 server for free. I've been using it for minecraft and it performs excellent, especially combined with PaperMC
> 2 AMD based Compute VMs with 1/8 OCPU* and 1 GB memory each.
> 4 Arm-based Ampere A1 cores and 24 GB of memory usable as one VM or up to 4 VMs.
> 2 Block Volumes Storage, 200 GB total.
> 10 GB Object Storage.
> 10 GB Archive Storage.
Just to compare with Hetzner - you would typically be able to get 16 or 32gb ram, an i7 with 4 cores and 2x1tb disk (no ssd). I'm guessing single core performance might be higher than the arm offering - and a better fit for minecraft.
Ed: although maybe core count would win out with PaperMC?
If the cheapest is $25/month, I think it would just make sense to get a used school/office PC. I was able to pick up a used Haswell i5, 8GB of memory PC for $200, which you would probably get better performance out of, and break even at ~8 month mark. That being said, pandemic prices might shift those numbers.
Nothing wrong with getting a used pc, but do you have 1gb uplink at home?
In the cheapest offers right now, there's a Intel Core i7-4770, 2x 2 TB Ent. HDD (spinning rust, not ssd) with
32 GB ram. And a xeon box with ecc ram (same, low price).
Also, even if you do have a 1 Gbps uplink, a lot of providers will throttle you if they suspect you're actually hosting a server that tries to utilize most of that bandwidth in anything but bursts.
It does run on arm64, I am currently running a forge server with a few mods on this exact setup. It's just for a few friends, I don't think we've had more than 5 or 6 on at a time but it seems to be performant enough.
It's not perfect but it ended up being a better experience than we were having with minecraft realms.
What were the specs of the server and which Spigot fork did you run?
A month ago I ran a Paper instance from a Digital Ocean instance with about 4 GB of RAM and OpenJ9 JVM and it never dipped below 20 TPS even with a larger render distance. This was on Vanilla 1.17.
Have you noticed much difference between a HotSpot-driven JRE and OpenJ9?
I am somewhat irrationally biased against J9 because they made us stick it in everything at IBM, but I'm willing to reconsider for better Minecraft performance.
I have yet to see hard data on it, but the folks that have been in the trenches doing client-side mods on Minecraft swear that J9 is the superior JVM for the latest versions of Minecraft (1.16+). I don't know that anyone has really done a proper benchmark.
> a small Linode [...] it struggles to keep up with just a half dozen of us
I run my server on a 32GB Ryzen5 PC with the world on an SSD and it often struggles with just 2 players no matter what options I tweak. Keep dreaming of the day the Java parts get the same performence as the Bedrock parts (but I know it'll likely never happen.)
Something wierd about a vanilla minecraft server is that sometimes less RAM is better. If you are using a good enough cpu going from 4 gb allocated to 2 gb allocated massively increases performance for low numbers of users because garbage collection runs more efficiently.
You can tune the garbage collection manually, but super counterintuitively this fixed most of my problems.
While developing this did you come up with any best practice recommendations for individual nodes?