Where’s the Memory?

    June 20, 2006

I had happened across How To Remotely Monitor Memory Usage and smiled when I saw this :

As one example of this phenomenon, the SUSE Linux system that I use for my public Web and e-mail server originally only had 512 MB of RAM. This should have been more than adequate for the lightweight load that I was expecting, but the system tools showed that the box was constantly running low on memory. And even after I bumped it up to a full gigabyte, the system still complained.

I had just gotten ready to add another gigabyte when I discovered that most of the memory was actually being consumed by SUSE’s aggressive disk-caching algorithm. Every time I added more memory, the operating system drank it up for the disk cache. In reality I was nowhere near hitting any kind of memory limit for the applications themselves, even with just 512 MB.

“That’s how Linux memory works”, I thought to myself, “he should have known that.”

Well, yeah, he probably should have. That once again points out the danger of assuming you know what’s going on based upon your knowledge of other systems. Being an expert at Solaris is helpful with Linux but it can also lead you astray. The opposite is just as true, of course.

But knowing something about “Linux” in general can also send you down the wrong path. I said that’s how Linux memory works, but that isn’t necessarily true either. The same article goes on to say:

RedHat and SUSE have substantially different algorithms when it comes to allocating cache memory. Where SUSE is very aggressive and tries to make full use of the available memory, RedHat leaves unused memory lying around in case it’s suddenly needed by something.

So if you’ve “learned” that Linux uses available memory for buffer caching, a RedHat system that doesn’t could confuse you. But wait – that’s not all.

“Swappiness” was introduced in 2.6 kernels. When some program wants more ram and the kernel has a bunch being used for caching, it could give up some of that. Or it could just make some other less active program swap out to disk. Which is a better idea? Well, of course that depends on what is going on right now. If there is heavy disk activity and programs that really aren’t being used, it would make sense to swap. If there isn’t much disk activity, or those swappable programs are apt to become active very soon, it makes sense to give back memory.

The idea was that the admin knows best what his system might require, so the admin should be able to say how much “swappiness” is desirable. If swappiness was set to zero (echo 0 >/proc/sys/vm/swappiness), the kernel would give up cache to anything that wanted ram; if it were set to 100 it would do the opposite and let the processes swap.

Well, great, but what if you don’t have a clue because sometimes your system has unused apps and sometimes it doesn’t.? No problem, somebody came up with something that would make a good guess on the fly: Autoregulated VM Swappiness Patch.

“[The] amount of swap space consumed is also taken into account, and the size of it compared to the physical ram is taken into consideration when making its effect on the value of swappiness. With this patch, this should make any machine that has swapspace as resistant to OOM as possible. This version by default autoregulates the swappiness, but also allows you to choose a manual setting if you so desire by echo 0 > /proc/sys/vm/autoswappiness and then setting the swappiness the manual way as previously. This makes comparison with autoregulation easy.”

But then: ( Bugzilla Bug 54560):

As of 2.6.7, gentoo-dev-sources no longer contains the Con Kolivas autoregulated swappiness patch. This is important for desktop machines, to avoid having large amounts of memory swapped out on prolonged system inactivity (causing an annoying swapstorm when the machine is used again).

Desktops and laptops are apt to sit unused for long periods. Therefore unused processes tend to get swapped. If the kernel were really smart, it would notice that *nothing* of importance is happening and just leave things as is, but apparently we haven’t reached that point yet. Or if we could have so much real physical ram that we could eliminate swapping entirely..

Well, the minute you say “eliminate swap”, someone will point out that you *need* swap. There’s a long discussion of swap and ram at Kernel trap: Is Swap Necessary?, but most of the conversation there is at cross-purposes. For multi-user machines, you definitely want swap because you want to be able to be truly interactive: moving one program aside so that another can run. But for a single user machine, you may not want to swap: it might be preferable to have the system refuse to start up a new app and let you decide what you’d rather get rid of to make room for it.

Or maybe not. I’m in favor of control, but the default probably should be to swap. Then again, maybe the default for desktops and laptops should be different.. it remains a complicated and difficult subject.

*Originally published at APLawrence.com

A.P. Lawrence provides SCO Unix and Linux consulting services http://www.pcunix.com