On Sat, Sep 15, 2018 at 11:55 AM, mcatanzaro@gnome.org wrote:
On Fri, Sep 14, 2018 at 10:43 PM, Chris Murphy lists@colorremedies.com wrote:
OOM killer might be worse if it's really arbitrary or non-configurable. What if it kills PackageKit, or gvfsd, or the journal? A sluggish system to the point it's unusable is bad, but it's probably less bad than services silently dying. Anyway, both are bad.
I suggest immediately powering off immediately before swapping is required. There will be horrible data loss, but it's better than the status quo (hanging until the user decides to manually power off regardless). We shouldn't be forcing users to decide for themselves when to press the power button.
It's a real shame.
At this point you're just repeating this without saying if you've tried either zram or zswap. In my recent experience, I can't reproduce what you're talking about in a VM with a hard drive backed swap. That's slow as molasses but it does work, whereas OOM killer is pathological especially for an installation environment.
So in terms of priorities:
1. Enable zram.service by default on live installs. They're enabled on netinstall. I filed this request 6 months ago. https://bugzilla.redhat.com/show_bug.cgi?id=1562278
2. Up the memory minimum from 1G to 3G or 4G on getfedora.org. Even 2G is unreliable without some kind of swap, zram or drive backed.
Following the more immediate issues, could someone who knows more about zram vs zswap than I do, propose a feature for Fedora 30? Drop traditional swap, and enhance with zram or zswap? I don't see how you can just drop swap entirely, it's not a good enough stop gap for solving whatever the bigger problem is.