I've bumped into misconception that swap is useless because computers have so much DRAM.

Not true. It is just kept relatively small (like 2GB and similar figures) because its main job is these days to provide depth in failure tolerance for near OOM situations.

Back in the day it was for "spare slow memory space", which is of course not applicable anymore.

Except: if hibernation is used it obviously need to be larger than the system memory.

@jarkko There are unfortunately lots of programs which do reserve memory (and a lot), but do not actually use it after that basically ever. It's much better to use the RAM for actually useful things than junk that's never needed or used after allocating it.

@sl This another totally misunderstood concept. Only thing that gets reserved is a range of addresses. Memory pages get reserved only when they are accessed.

@sl Look into "reserved set size".

@jarkko

Old school classic RAM example

First, I run a program that roughly allocates and fills 3 GB of RAM.

RSS:
python3 3156956

Then, I run a second instance of it. Wait for a while, and:
python3 1044036
python3 2380900

Neither of the programs has released any RAM. But the RSS is now roughly half of the RAM that the programs are using. But why? Well, that’s just because the rest of the stuff is in swap and not included in RSS anymore.

I’ve found out that memory is one of the most misunderstood concepts with modern operating systems. And that’s probably because it’s a very complex topic. I really don’t know anything about it, but over the decades, I’ve tried to get even the most important stuff right in my mind, because I’ve dealt a lot with VMs that had very limited RAM and high memory loading at least for short periods of time.

I often hear sysadmins, DBAs, and users claim that the system is just running out of memory, and they’re panicking and asking to get more memory before the system crashes. I just log in, run a program that allocates 4 gigs of RAM, fills it, and releases it. And whoa, now we’ve got 4 gigs of free RAM on a system that was supposedly out of RAM?!

So, when the system is out of memory, just run this to get 4 gigs of more RAM and reduce the RSS of other programs. (Yes, I used the 3 GB version for this demo, but I’ve bumped it up to 4 GB.)

py -c "grab_it = b' ' * 4 * 1024 * 1024 * 1024"

Of course you can use whatever language or memory reservation scheme you like instead of Python.

And for sure, I do realize that you’re posting from kernel.org, which scares me. What’s the most important thing I’ve missed in my post and got completely wrong? The VMs I’ve dealt with are mostly Windows, but basically behave quite similarly. - A link is also a valid reference. - Thank you.

@sl @jarkko you are correct, RSS stands for Resident, not Reserved, and behaves as you say. What Jarkko means as "reserved" are the virtual memory areas, reported as VSS or VSZ and includes also areas that were allocated by mmap() without MAP_POPULATE and not yet accessed.

@sl @jarkko and yes programs that will populate their areas so they start using actual memory, but then don't touch them anymore, are wasting the memory and it could be just swapped out.

@vbabka @sl My favorite call is fallocate() with FL_PUNCH_HOLE flag ;-) With just fallocate() and mmap() fairly complex memory management can be implemented.

Everyone’s favorite gpg-agent demonstrates pretty well vsz/rss ratio:

 main 21s
❯ ps -h -p "`pidof gpg-agent`" -o pid,comm,vsz,rss
  PID COMM           VSZ    RSS
 3739 gpg-agent 410782624   2480

For the sake of example this was executed in macOS (and for the record, numbers are KiB).

@vbabka @sl If I created a new systems language from scratch I’m not sure if I included to its stdlib anything else than memory mapping primitives.

I don’t understand why even latest of latest languages still implement POSIX API’s “best of tape drives” parts. Like e.g. Rust has only the tape drive API for files and you need to use external crate calle rust-vmm/virtual-memory to get mmap 🤷