Does it actually work?

Posted Apr 10, 2025 12:28 UTC (Thu) by khim (subscriber, #9252)
In reply to: Does it actually work? by intelfx
Parent article: Three ways to rework the swap subsystem

Why is it non-goal, BTW? I know that Yandex was using FreeBSD for years because it could save money that way: with two sets of daemons on the same machine and 99%/1% split everything was perfectly behaving – the 99% side was perfectly responsive and could handle the traffic without being swapped out while 1% would get small percent of that same traffic while responding slowly because it would be trashing like hell (but responses would go to log and never to a human thus it was Ok).

Linux could never pull tricks like these… but I have no idea why. Because they were declared “non-goals” by someone?

P.S. Eventually the had to adopt a different way of doing experiments and switched to Linux… but mostly because it was becoming harder and harder to keep FreeBSD going when most hardware favors Linux.

to post comments

Does it actually work?

Posted Apr 10, 2025 13:32 UTC (Thu) by intelfx (subscriber, #130118) [Link] (1 responses)

It's a non-goal for the swap subsystem (in the sense that no amount of making swap faster will make it as fast as RAM).

It may or may not be a goal for the overall memory management strategy; limiting thrashing to some parts of the system should be somewhat doable with suitable application of cgroups or perhaps other, more auto-magic technologies (like aforementioned MGLRU working set protection). But ultimately, all that does is constrain unusability to those parts of the system that *are* thrashing (and likely degrade the overall throughput). Making *those* parts of the system usable is still a non-goal because it is fundamentally unachievable.

Does it actually work?

Posted Apr 10, 2025 13:39 UTC (Thu) by khim (subscriber, #9252) [Link]

> Making *those* parts of the system usable is still a non-goal because it is fundamentally unachievable.

Why not? MacOS and Windows achieve that… somehow.

Sure, when it takes 50 minutes to compile something that can be compiled in 5 minutes without trashing… it's not good, absolutely – but it's still usable.

When the same thing takes days (not sure how many days, though, I stopped the experiment after two)… it's not usable.

You may say that it's not worth it (Apple did all that work, apparently, to claim that MBA with just 8GiB of RAM is as good as any other laptop with 16GiB of RAM), but that's something radically different from “fundamentally unachievable”.