How Big Should My Swap Partition Be?

For the last 10 years, I have been asking people more knowledgeable than I, "How big should my swap be?" and the answer has always been "Just set it to twice your RAM and forget about it." In the old days, it wasn't much to think about — 128 megs of RAM means 256 megs of swap. Now that I have 4 gigs of RAM in my laptop, I find myself wondering, "Is 8 gigs of swap really necessary?" How much swap does the average desktop user really need? Does the whole "twice your RAM" rule still apply? If so, for how much longer will it likely apply? Or will it always apply? Or have I been consistently misinformed over the last 10 years?

Comments

What Has Changed?

[eldavojohn]

'Is 8 gigs of swap really necessary?'

With a 750GB hard drive selling under $100, what has changed?

Yeah, your 256MB of space was trivial when you had a 30GB hard drive ... and 8GB of space is still trivial with a 750GB hard drive.

That said, I'll forward you some common information on paging.

Linux and other Unix-like operating systems use the term "swap" to describe both the act of moving memory pages between RAM and disk, and the region of a disk the pages are stored on. It is common to use a whole partition of a hard disk for swapping. However, with the 2.6 Linux kernel, swap files are just as fast as swap partitions, although Red Hat recommends using a swap partition. The administrative flexibility of swap files outweighs that of partitions; since modern high capacity hard drives can remap physical sectors, no partition is guaranteed to be contiguous.

I'm no expert but the short answer to this is to look at your swap partition as your extended virtual memory. By saying that your swap partition should be 2x your main memory is like saying that you will never use 3x of what your main memory is (in this case 12GB). While that rule of thumb is a good one, there may in fact be applications today in the graphics and processing world that require insane amounts of memory. While Firefox is probably never going to reach that critical mass (nor will most average programs) it's probable that a few years from now it will be common place. I know it's insane to think of but 'ought to be enough for anybody' is not the phrase you want to throw around in the digital information world.

It's those days when I'm playing Warcraft through wine, listening to streaming radio through Amarok, have 20 windows open behind it, idling a LAMP server for my development projects, running a vent client, some form of news aggregater, pidgin & an e-mail client hooked up to several POP3/IMAP accounts that I am happy I erred on the side of a whole ton of swap space.

Re:What Has Changed?

[orzetto]

I have an Eee 901 [...] I decided to be bold and installed Hardy with no swap partition.

There are better reasons than boldness for not using swap on an Eee. They use solid-state drives (except some 1000-series models and the 904), which are faster than mechanical devices but can be rewritten fewer times. To make sure your drives last longer, do the following:

1. Mount partitions with noatime, or relatime if you are using one of the very rare programs that use atime (mutt is the only one I know of);
2. No swap partition, which would predictably have many more writes than the rest of the disk;
3. Mount /tmp on tmpfs so that temporary files do not wear the disk.

Sure, without swap and with tmpfs you will have less memory available, but I have an Eee 900A and I bought it as a presentation machine, possibly for some occasional work while travelling, not as a workhorse.

Re:What Has Changed?

[kasperd]

what I do is get the OOM killer to take action and kill some processes sooner.

Just killing processes more or less at random when the system runs low on memory is not a good idea. (I know it is not completely random, but there surely ain't any guarantee that it will make sane decisions). What you really want is for programs to get an out of memory error when trying to allocate memory, and then they can shut down as gracefully as possible. (Would be neat if the choice of who get the first ENOMEM was chosen by the heuristics that would otherwise kill the process, but I guess that has not been implemented). Guaranteeing that you will never need to kill a process because you are out of memory means the kernel will have to not commit to more than can be backed by memory and swap. However since actual usage tends to be somewhat lower than what is actually committed to, that would be a bit wasteful. This is the main reason why it makes sense to have a large swap partition that is mostly unused. Just so you have backing for the amount you need to commit to in order to optimally use the physically available memory.

You typically wouldn't want to make use of most of that swap. So once any significant amount start getting used, you'd want to start giving ENOMEM errors. And that should help ensuring that the swap will only be used for a short time. There are a few pieces of data in virtual memory that are only used under very rare circumstances, and those it is nice to have on swap so they don't take up precious memory. So the aim is not to have zero swap in use, just some low number of pages that are really not needed in memory.

Is there any kernel out there that gets all of this right? I don't know. But at least those I know about can be tweaked to do pretty well.

Re:What Has Changed?

[rcw-home]

I am thinking of reducing the amount of swap on my primary compute server

For best performance, don't reduce your swap below the amount of RAM you have, unless you want to get rid of it entirely. The reason is that Linux 2.2.x and later will, when your disks are idle, preemptively copy your physical memory to swap - that way if you do run out of RAM, all Linux has to do is reuse that RAM for other things - your application's virtual memory has already been written out to disk. This can't work as well if the swap space isn't there for it.

With 2.0.x and earlier, I would have recommended you pick the amount of virtual memory you think you need, subtract the amount of physical memory you have, and set up that much swap. With 2.2.x and later, I recommend you pick the amount of virtual memory you think you need, and set up that much swap.

For what it's worth, Windows NT derivatives do the same thing.

Re:What Has Changed?

[TXG1112]

Base 2. Storage numbers using base 10 are for disk manufacturers that are filthy liars.

Re:What Has Changed?

[duguk]

I guess I'm one of the few people who uses windbg to do this then?

Try it out, it's amazingly useful for debugging BSODs.

Re:What Has Changed?

[Mr Z]

It's base 2 when dealing with RAM and base 10 when dealing with disks. Pretty consistently. It's generally base 2 when dealing with throughput.

Pop quiz:

Throughput: How many bits per second peak can a 14.4kbps modem move? 1.544Mbps T1 line? 10Mbit Ethernet?

Disks: How many bytes are on a 1.44MB floppy? A 2.88MB floppy? A 650MB CD-ROM?

Answers:

Throughput: 14,400. 1,544,000. 10,000,000. Hmmm... so much for base-2 throughput numbers. And yet, when you see the "kB/sec" rate in your browser download dialog, that is most likely in a 1024 byte/sec quantities.

Disks: 1,473,560 (1440 * 1024, a mixture of base-10 and base-2), 2,949,120 (2880 * 1024, again a mixture), and 681,984,000 (purely base-2, derived from 333,000 sectors * 2048 bytes/sector / 1,048,576). And yet when you look at disk capacities from most computer software, it's reported as purely base-2 sizes.

So, what's consistent about this again? RAM seems to be the only thing that gets it right most of the time, though I do remember seeing plenty of adverts for Commodore 64s that listed them with 65K of memory back in the day.

And for the real brain bender: If we agree that bits should always use power-of-2 meanings and everything else should use power-of-10, what do we do when the two collide, such as when talking about areal density? (That's bits per square meter.)

That said, whoever came up with the names gibibytes, mebibytes and kibibytes must have wanted us all to sound like we have a speech impediment or something, as the pronunciation for these sounds worse than baby talk. I'll stick to saying gigabytes, megabytes and kilobytes and their understood power-of-2 meanings where it makes sense, knowing full well that it has deep flaws. It's just an unfortunate circumstance, but most of the time it thankfully doesn't matter.

--Joe

Re:What Has Changed?

[piojo]

Linux will use swap sometimes even if you don't fill up your RAM. It can swap out idle programs and use the recovered RAM for file caching which gives a performance boost to the file system.

Conversely, if you have enough ram for file caching and running applications, then you will get a performance boost from disabling swap (because applications would be faster if they weren't ever swapped out).

Definitely not twice...

[42forty-two42]

The origin of the 'twice real RAM' came in the early days of windows, in which windows could not use any swap unless you had at least as much as real RAM. That's been gone for ages now - and you should actively avoid too much swap.

If you allocate, say, 8G of swap for 4G of RAM, most of the time almost all of it will go unused. If it actually /is/ used, your machine has probably spent the past hour or so frantically swapping to try to accomidate this 12G request; ie, your system is completely unresponsive due to every program being mostly swapped out. The additional swap merely delays the out of memory event, and in the meantime you can't control the machine.

Swap is still useful for holding data that's not part of the working set, in order to free memory for cache; but this shouldn't be very much RAM (256-512mb should be enough). It's also useful for software suspend on linux - if you have a laptop, make it a little bit larger than physical RAM. And always have /some/ - linux's memory manager doesn't like having none.

What Oracle Wants

[stoolpigeon]

If you were running Oracle - here is what they recommend:
    RAM -> Swap Space

    1 GB - 2 GB -> 1.5 times the size of RAM
    2 GB - 8 GB -> Equal to the size of RAM
    more than 8GB -> 0.75 times the size of RAM

I don't know if this would carry across to general computing - it seems to me if it's enough for an Oracle RDBMS server, it ought to do it for most things.

More sensible suggestion...

[Vexler]

Reading through OpenBSD's FAQ:

"The 'b' partition of your root drive automatically becomes your system swap partition. Many people follow an old rule of thumb that your swap partition should be twice the size of your main system RAM. This rule is nonsense. On a modern system, that's a LOT of swap, most people prefer that their systems never swap. You don't want your system to ever run out of RAM+swap, but you usually would rather have enough RAM in the system so it doesn't need to swap. If you are using a flash device for disk, you probably want no swap partition at all. Use what is appropriate for your needs. If you guess wrong, you can add another swap partition in /etc/fstab or swap to a file later."

HTH.

Just test?

[rasteri]

Just make a note of your virtual memory use every hour or so (or just whenever you remember) for a few days/weeks. Then just give yourself maybe 2-3 times the peak usage.

I imagine different people will need different amounts of swap space, so use a size that's right for you.

If you need crashdumps, same rule applies

[bugg]

If you're debugging your kernel or are helping people to debug your kernel, and are generating crashdumps either manually or as a result of kernel panic, you need your swap to be twice as big as the memory so it all fits comfortably (You can probably get away with X times bigger, where 1X2, but 2 is a safe number).

To my understanding that's always been the reason for the rule of thumb about doubling the memory. If you can afford the disk, go for it, because you never know when you might hit a panic and need crashdumps. If you are in a live environment and are sure you will never, ever need or even want crash dumps, and the disk space is at a premium, you can size it based on need.

Another thing to keep in mind is that as you have more ram, you have more pages, and the whole point of swap is to get pages to disk as well in case you need to free up physical ram quickly.

It Depends, but at least as much as RAM

[rcoyner]

In the end, it depends. If you are running several memory intensive applications you're going to want more swap space. At the very least, you should have as much as your RAM because when you hibernate it takes all the pages in your RAM and puts it into the swap space before powering off your computer. I wrote about this a while ago: http://www.bytetrap.com/blog/2008/06/02/swap-space-linux/

Re:It Depends, but at least as much as RAM

[Anonymous Coward]

Yeah, hibernate actually DOES mean "use no electricity". Perhaps you mean "suspend"?

Oh, nonsense

[Anonymous Coward]

2X RAM was the standard rule of thumb at Sun, for SunOS long before Windows was around.

If anything, Microsoft ripped it off from Sun.

Re:Oh, nonsense

[Caduceus1]

Yes, I believe it was the BSD memory manager (possibly earlier, V7 maybe) that had the 2xRAM rule. Less and you could have issues - more was wasted disk space.

BSD was the foundation for SunOS (pre-Solaris 2.X), Ultrix, etc. so they all inherited this requirement - and from there the "requirement" became gospel on other systems.

I've actually never heard the 2xRAM in relation to Windows, but it certainly predates it. I was setting up systems with 2xRAM when Windows was a DOS app... :)

Re:None

[hey!]

Well, I do occasionally need more than 2GB of RAM, without there being a memory leak. I've been running GIS programs, an IDE, a couple of RDBMSs, and then I fire up the old compression program...

Which brings me to my point. The question "how much swap do I need" is probably meaningless, even for a given amount of memory. There are people who find 2GB with no swap fine, and others, like me, who probably could get by with 2GB of RAM and maybe 512MB of swap, and others who might need more.

I think the 2x RAM rule of thumb has one virtue: excepting certain exotic kinds of systems, it's fairly safe that anybody who finds themselves needing more than that is probably feeling a world of pain that can only be fixed by getting more RAM. On the other hand, in most cases 2x RAM amounts to a trivial amount of disk. Probably most people could get by with 25% of RAM, but the value of thinking about whether that is true for you is very likely less than the cost of the disk space.

Common sense applies. If you have some kind of scientific computing device with a gazillion bytes of RAM, your swap requirements might not be related to your maximum RAM requirements at all. If you're running some kind of operating system that launches a bunch of rarely used garbage, you probably ought to think about your swap. I had awful problems with Vista until I figured out the page file Windows created had something like eight thousand fragments. I was actually better off getting rid of the page file

Errr... check your math.

[sirwired]

8GB swap on a 120GB drive is 7%, not .07%. On a 200GB drive, it's 4%, not .04%, etc.

SirWired

Re:Here's how big

[BitZtream]

So I'm a FreeBSD guy rather than linux, but I'm going to assume that Linux also supports 'limits' that define the maximum a program can utilize before its denied access to more resources. You won't get a normal app on my FreeBSD boxes to use more than 256M of ram, they aren't allowed. There are 2 exceptions, the PostgreSQL server on one of the machines, and the bot that connects to that database. They both deal with large datasets on regular basis so they are allowed to use more ram. Now mind you, these machines are used for my personal development projects and they aren't really 'servers' in the sense that they see real load. My instances of apache don't NEED a lot of ram, some do.

My point is that there are other protections in place that prevent an app from 'running away' and taking a properly configured machine down.

Second, swap can be VERY useful even if you NEVER run out of ram. The OS can swap apps that have used memory but aren't actually doing anything with it out, and leave that memory available for file/disk caching, which can make performance FAR better than if you kept the idling apps in memory and had less available disk cache. Some apps avoid buffering things in memory because its both easier and most times more efficient to use the disk and let the OS manange the buffering. I've seen NT based OSes aggressively swap out things that aren't in use just so there is more memory available for disk cache, and it makes sense cause there is a lot of crap the kernel and other apps load up that is very RARELY needed, if ever.

So while you can ( and did ) point out the potential pitfalls of using swap, your examples don't apply to any modern OS. I'm excluding Windows from that statement cause lets face it, its not exactly modern at the core. Modern kernels are FAR better at deciding what to swap than you are in almost every case, just like compilers can do a far better job of optimizing applications that most developers can. Yes some can do better, but its not likely you are, and certainly not the guy asking this question.

In short, if you're going to try to get technical with why you wouldn't want to use swap, at least use examples problems that weren't solved years ago.

And for reference, you configure your swap poorly if you do what you say.