Hi,
Do you have suggestions for kernel tweaks for getting the most out of a RAM limited system?
I am running a service requiring 2GB of RAM (netbird) on a VPS which has just 1 GB of memory. I am doing so because I am a stingy bastard and I use only free VPSs for my personal use so I get what I am paying for.
Because of this hardware limit in about 12 hours from service start I begin swapping a bit too much. This would still be manageable but soon the hypervisor gets really pissed and steals up to 90% of the CPU. So the only solution so far is restarting the docker containers every 12 hours (not great, not terrible).
Looking to improve this, Iam now experimenting with ZRAM and swappiness and it seems some benefit can be achieved by using some of the Linux kernel feaures. Is there anything else I should look into?
Thanks for sharing your experience, indeed the distribution is relevant here. I am running Arch (BTW) on this VPS which idles at about 300 MB with dockerd and containerd, I am not sure how does exactly compare to Debian on RAM usage (I have a couple of other VPS running debian which seem to use a little bit more RAM but it could be because those images are bastardized by the addition of cloud provider services). In any case my setup is pretty minimal, to get some large benefit there I fear I should use something without systemd :/
Yeah the 300meg isn’t going to get much less. Switching to Debian won’t change much there. Perhaps you can look into running a minimalist container distro if you are just using the machine for that. I personally want to check out Talos, there’s also RKE and Burmilla. No experience with them, to me the memory doesn’t matter much because I run a homelab. So I currently just run Debian and k3s. On my systems the containers are actually what gobbles up all the memory. If you’re using public container images, there’s a good chance the memory configuration on them isn’t optimal. Especially JVM services are a lot of the time configured to just use whatever is available. If you give them less memory they will do more garbage collection. So if CPU is less an issue then mem, that could be worth looking into (it’s just parameters you can pass on startup). Hopefully any of this is of use. Good luck :)
Thanks again, I will look into your suggestions, never heard of neither Talos, RKE or Burmila. Indeed I should also look if I can do in the containers. The problem is only this database “CockroachDB” which is extremely memory hungry, maybe I can change something there.
[This comment has been deleted by an automated system]
Thanks, this is a really good point, I can try to replace the identity provider! I did not realized that cockroachDB was only a Zitadel requirement! There are many great alternatives for mesh VPNs, netmaker, nebula, and headscale as you mentioned and all of them are much lighter. I ended up hosting netbird as it is natively able to traverse my corporate NAT (maybe headscale could do it as well, I did not try it since I do not like having to configure registry keys on windows clients and losing the kernel wireguard speed on linux clients) .
CockroachDB is a clustered version of PostgreSQL you probably should be able to replace it with that. But running a full RDBMS with the resources you gave is not great. SQLite would be a better fit for the resources available if the tools you run support it.
One of the answers in this discussion made me realize that this database is required only by the identity provider which I can change from the default. Considering this is a far less essential dependency than I thought I will get rid of it completely.
Sounds like an excellent suggestion.
If you’re idling at 300mb with containerd running, you’re not getting better than that with a modern general-purpose distro. As others have said, switching to another vps’ free tier that offers more is your single best bet by a mile. About the only options you have on this one are compressed ramdisks being used as swapfile (zram) and literal disk swapfile to get you the rest. It’ll be very slow though if you have to load half your workload on shared platter swapfile.