Databases Without an OS? Meet QuinineHM and the New Generation of Data Software
dataware.devhttps://dl.acm.org/doi/pdf/10.1145/3709714
sounds pretty similar motivation-wise to this. would be interested if linux as a host system is still needed. and if not: how is heterogenous hardware supported without virtualization?
We don't think it is needed. Currently, the only thing which is stopping it from running on bare-metal no hypervisor are some drivers for virtual devices (eg. virtio net driver).
Challenge is, ofc, supporting wide range of hardware options, but we don't intend to. We plan to support only modern x86_64 CPUs (arm is on the roadmap) and a select few NICs.
Current implementation is written to be hosted on public cloud, so we implement what is needed for that, but in the future we intend to allow self hosted version, but only on selected server hardware.
Now, if we have "big" server with like 256 cpus, etc. we can still split it many smaller "vms". Since the only code that runs with Hardware Manager is code we write, we can just add config options which pin execution to resources, eg. ram, we can just allow it to use from 4gb-8gb phys addr. Shared hardware like NIC is a bit trickier, but we do have some tricks up our sleeve which allow sharing.
> which is stopping it from running on bare-metal no hypervisor are some drivers for virtual devices (eg. virtio net driver).
why would you need to support virtual devices, if you are not in a virtualized environment?
Yea poor phrasing in that sentance.
All code we have can already run on bare-metal except some drivers, since most are only for virtualized devices.
Hopefuly that clears what we wanted to say.
If we implement eg nic driver for physical device then we can run bare-metal no host at all.
Sounds great! This is the future!
Glad you like it
Btw, we fixed some whitelist issues and if you want to play around with it, database creation should work now (if not with existing account you can create new one)
Personally iam skeptical about the performance. If possible, i would like to see any published report of benchmarks eith 2000% performance gains.
We are actively working on it (also writing a paper). Performace gains like this aren’t unexpected for systems like this see: https://www.cs.cit.tum.de/fileadmin/w00cfj/dis/papers/cumulu...
That type of tech + redis being quite bad implementation allows those gains (up to 2000%, we even have commands showing ~2200%, mostly lists since they have very cache unfriendly implementation)
Wish if you also add about team, ompany and other details to look it more real.
Will add it as soon as possible, thanks for feedback.
We are 3 students, founded in 2024. We have experience at Microsoft and Databricks.
The more I'm thinking about it, the more makes sense to go in this direction. I think you might be onto something. Wish you luck!
Site is bare of real info, page links just lead back to home. Please update the site with content or fix the links, could be interesting.
Hi, we fixed the main link now (hopefully it works now, thanks for pointing it out).
This is just initial version of the website, we are working on a new one.
Glad you find it interesting :)
If you have some questions feel free to ask.
1-1 mapping with physical ram? so your in-memory data actually has physical addresses?
Yes, since we never run untrusted code, there is no reason to use more complex mechanism. Memory is 1-1 mapped with physical ram. This also allows us to control cache even better (side note, we use huge 1gb pages for memory mapping, to ease pressure on TLB/MMU).
The world’s first Hardware Manager, built to run databases at bare-metal speed.