Long story short, I need to find out how much memory I have in use. Yeah, I know, free(1) is your friend. However, there's a bug with the particular virtualization I'm working with, and so inside the container, that isn't always reliable.

At the moment it is...and I ran ps aux, but adding up VSZ or RSS or VSZ-RSS isn't right.

So what do I want to look at in ps to find out how much memory each process is using so I can add those up?

Maybe just have a look at /proc/meminfo ?.

free(1) just reads meminfo - it's really the same thing.

It is, frankly, extremely unlikely that there is "a bug in the virtualization you are working with." You are probably mis-reading the situation.

Aside from the fact that memory allocation changes from one microsecond to the next, much of the memory that you might be measuring is a shared resource, and the entire allocation and management scheme is "lazy." In other words, you might be counting the same bytes twice, and, if there is no actual, continuous pressure being applied to the memory manager, the system (by design...) is not making any efforts to economize.

Remember, the system has every incentive not to "clean up" something that, a few microseconds from now, someone might need to use again.

It's a bug with OpenVZ interacting with the .32 kernel on the host. To quote my provider:

"My apologies, this is a bug due in the .32 kernel the node is running. The only thing affected is how RAM is displayed, so it's not major enough to warrant a full node restart until we're ready for another round of upgrade."

I was hoping to get the same resolution as free(1).

Never going to happen I'd reckon.
See the "ps" manpage comments re size and RSS. Everything has page tables, even kernel threads - and if you happen to run something like Oracle, they can add up quickly.
Then there's things like slab. There's just too much you can't see, let alone how to count what you can see, like shared code. smaps tries to help, but you probably can't get the reference count to make a sensible total for the entire system.

1 members found this post helpful.

The bottom line when using commands like that is that you are in it, hurtling along the road along with everyone else who's on the same penguin-bus with you. You don't have an "outsider's" perspective, and you can never get one.

Each of the numbers produced by these various commands are intended only for comparative instantaneous inspection within their own current output ... and even then, "only approximately." They are not intended to "add up."

1 members found this post helpful.

OK, thanks everyone for the replies.