Ask HN: I have 24 core server with 1TB of DDR4 RAM, what should I run?

I would definitely run KVM, and set up virt-manager locally and connect to the server remotely over SSH. Whether you sell VPSs to friends or not, you can use it as a quick scratch pad for spinning up different os's for various testing purposes, both long lived and short.

I have a similar piece of hardware but 256 GB instead of a terabyte of RAM, and that is what I do. It has come in incredibly convenient to be able to spin up VMs as needed. I started creating different VMs for purposes. I would have normally just used the same host for, and have really enjoyed it.

I also run about a dozen personal services on there, such as audio bookshelf, archive box, jellyfin, navidrome, and more. Surprisingly, the archive box instance uses quite a bit of memory and CPU, so the box does get a fair amount of exercise. I have not looked very closely, but I believe archive box is using that memory and compute mainly for running headless Chrome. I have a browser extension installed that archives nearly every page I visit automatically, so especially during busy browsing times, I keep that thing running pretty hot.

In the past I set up a self-hosted openshift instance on it, spread across six VMs. I actually loved that and the only reason I'm no longer running it is because I broke it by messing around with risky and dangerous things that no sane person should ever do, and then did not want to dedicate the time to rebuild it. Someday I will recreate it again.

Whatever you decide to do with it, this is a really awesome problem to have!

Chrome with 12 tabs open

Run a rails server, a DB and you'll probably not need to worry about scale until you're earning millions a year.

Here's a random idea: run a freely available electromagnetic simulation agent! My favorite EM solver has an "agent" that can run on a server, you simply point your local GUI to the agent and take advantage of more computational power.

https://www.simberian.com/

I have a similar setup, but with 20 cores and only 144gb ram but a 4070.

I run proxmox on it, have servers for pihole and networking, AI, docker & portainer & jellyfin.

I've just about run out of the things I want to do with it and I barely use 10% of its capacity.

You could run hundreds of nodes for the Autonomi peer-to-peer distributed storage network, see https://autonomi.com/ .

Simulations of maximum concurrency.

With hardware like that I would research a couple of things:

* Maximum number of servers, HTTP/WS, I could run simultaneously. I am working on a server application to do this right now. Each server should be 1, 2, or 4 ports. You could run HTTP over WS which allows both protocols over a single port. If you want to allow both TLS and insecure connections it would still be two 2 ports, or 4 ports if you are isolating HTTP and WS from each other.

* Maximum number of simultaneous sockets. I would test for the maximum number of open sockets connected to a single server instance and the maximum average number of sockets open to the maximum number of servers from the prior bullet point.

* Once you have confidence with both prior points I would then research the maximum amount of cross-talk. If your multiple servers can talk between each other then your servers almost achieve SMP. They could talk to each other via sockets like they are talking to everything else, but IPC communication would be even faster.

* Once you have all that then you have sufficient infrastructure in place to investigate more precise performance concerns. For example I have found that in my own personal implementation I could transmit messages via WebSockets almost 11x faster than I could receive the messages, but it could be that my own implementation is poorly executed. I also found that under the most ideal conditions my, likely poor, WebSocket implementation was still at least 8x faster than HTTP can go to over 80x and beyond performance after accounting for scale of high message frequency and socket concurrency.

* Once you have your performance bottlenecks identified you can then research performance bottlenecks on data transfer from a large data source of high frequency access. Then with that identified you can train AI on it for concurrency simulations that self-learns.

In the end you can sell your research. Consider that data centers require their own power plants to operate. While the hardware in those data centers is likely assembled in an efficient manner the software servers running on that hardware is often not efficient, and that costs millions of dollars a month just in electricity.

I’ve a proposal. I’ve created a Hypervisor that can launch lightweight and efficient VMs, similar to containers. Workloads in these VMs behave like isolated processes (isolated from other VMs) You can pack and sell 10x more VMs. Unlike traditional VMs, you don’t have to worry about turning them off so as to free resources. Would love to work together with you if you’re interested in using it.

For research, you can do things like PSO or Genetic algorithm optimizations. Those benefit more from CPU than GPU cause there is not so much matrix math going on, especially for complex fitness functions that require sequences.

You could do this with the small Llama model, where the fitness function is basically the ability generate correct code and self detect errors, and adjust the weights based on the optimization algorithm.

Maybe donate part of it to the NixOS hydra build farm? (not sure how easy that is). You can compile everything in RAM like that.

I know that my colleague runs proxmox on this, you can do many things with it, my company is currently running proxmox and set up many vms as kubernetes nodes, I always appreciate people who set up their own lab, that colleague of mine somehow has enterprise grade storage server, nics... and many more inside his house just to have fun

Maybe you can scratch the experiment and downscale it? Seems like you are very very very far away from needing something as beefy as that. Unless you got it for very cheap, I'd feel bad having so much money in idle hardware

Fuzzers, caches, graph algorithms, climate models, massive compilations, in-mem filesystems, VMs, Llama 3 405 B (will be slow)

I've been thinking about doing the same, seems much more cost effective. How much will you sell the VDS for?

Test rowhammer attacks to gauge how resistant/susceptible your dense memory sticks are vs. other systems.

Do some AI research that takes advantage of those characteristics.

Erlang?

> Ask HN: I have 24 core server with 1TB of DDR4 RAM, what should I run?

Xbill.

What is the framerate in Q3A, is the only benchmark

Some databases?

Make it a kubernetes node?

Make it run long async jobs like ETL kind of stuff.

Doom and Far Cry are the two extremes I can think of.

Try llama.cpp with the biggest LLM you can find.

An electron app (maybe two)

Honestly, that is a crazy server to put in a datacenter just for fun lol, and i have no idea what you could do with it, selling VDS space sounds like a good idea, especially if you can find someone who wants heaps of memory. Though I'm curious if you dont mind answering a few of my questions: - Where did you colocate (ie what city), do you live there? - How much was it to colocate (you can give ballpark if you dont want to give the full amount) / do you pay monthly or something, did you pay per U or is it something else - What costs do you pay (eg electricity or bandwidth that you use) Thanks, and good luck xd

Dokku

crysis