I have noticed that my home server is strangely using lots of swap (~5 GB), despite having only a few lightweight processes running and loads of RAM installed (32 GB).
Upon configuring Grafana + Prometheus, I noticed a trend where cache + buffer will progressively increase until swap starts to be used. My system and services combined will use ~8 GB RAM. Upon rebooting, the cache + buffer will start anywhere from 3--10 GB, progressively ramp up to ~25 GB in 1--2h, where swap will start to be needed (~3 GB). See the image attached for reference.
My swap filesystem is on an expensive (to me) SSD, and I would like to reduce its wear by as much as possible. I understand that swap can introduce only minimal wear on SSDs depending on its nature and that it can be harmless, but I am still not sure what is causing this behavior (and why) and whether I should worry about it or not. So I figured I should investigate what is happening here.
My main question is, how can I figure out what is causing this behavior? Is it expected? I am looking for guidance from others who are more experienced than me in the topic.
A little bit about my system:
I am running Debian 12 on an NVMe SSD containing the root partition (btrfs) and docker services. I also have two HDDs, one with persistent data (ext4), and the other with backups (ext4). This is majoritarily a single-user machine. I tried using the following kernel parameters, but it hasn't helped:
vm.swappiness=10
vm.vfs_cache_pressure=200
My docker services are:
- *arr stack
- jellyfin
- nextcloud
- immich
- open-webui + ollama
- pi-hole
- invidious
- romm
- nginx proxy manager
- grafana + prometheus
- other minor services that I don't think are doing much (uptime-kuma, stirlingpdf, vaultwarden, etc)
As a suggestion, ditch ollama and setup llama.cpp. it will work fine with openwebui and it's much more efficient. (Unrelated to the ram/swap issue)
Thanks for the tip. I have been looking into commenting out the ollama portion of the
docker-compose.ymlfile and setting up llama.cpp instead for some time now. I suspect I don't usually use features such asollama pullto justify a fully-featured application.Here is my experience https://wiki.gardiol.org/7-services/llamacpp
No ads no monetization just a wiki I keep for my future reference.
Never been happier than moving to llamacpp from ollama.
Can you talk more about this? What are you basing this on?
Llama CPP can run models offloading with CPU (es MoE models), you have much more control over how you run your models, and overall it's very much actively developed.
For starters and people without too much willingness to mess up with stuff, ollama is a great choice. Llama.cpp gives you that extra power and flexibility that is so much worth for people who like to tweak and do more.
My personal opinion, of course. But based on having used both and ditched ollama for llama.cpp, so I am also biased, keep in mind.
But I will hardly go back to ollama now :)