I frequently hear from people that they “don’t need swap” because their computer has a large amount of memory (nowadays, 4/8 GiB or even more). To a certain extent, this is true. However, it is still important to have swap space available in most situations.
It’s true that few desktop applications require anywhere near the full capacity of a typical modern PC’s memory, and even the complete set of programs someone might be running at a given time is unlikely to use the whole of available memory. Keeping that in mind, it often doesn’t make sense to allocate disk space that will seemingly never be needed as virtual memory.
On the other hand, not having any place for your kernel to hold dirty pages when memory pressure increases can severely degrade performance. I present here a case for why you should always have some swap available, and offer some advice on how best to allocate it.
[I’m targeting Linux systems here, but this advice should apply to other operating systems with little modification.]
Most power users are aware of the huge latency difference between physical memory and typical spinning disks1, and thus understand why allowing the system to swap virtual memory pages out to disk2 can be harmful to performance.
[If you don’t understand how virtual memory works and ignored that footnote in the previous sentence, read that then come back.]
What many users seem to ignore is that the operating system developers also know that there’s a large penalty involved with swapping data out to disk, so they will design the system to only swap when it is advantageous. If your system is swapping a lot (such as if there’s a lot of disk activity when you restore a previously minimized application), the situation can often be improved just by adding more memory.
Even when there’s plenty of memory available, however, it’s possible to see a large perfomance hit even when there is ostensibly a large amount of free memory.
All modern operating systems are aware of the large penalties involved with accessing non-volatile storage (anything that isn’t in RAM), so they keep disk caches in memory. On Linux, the disk caches are held in the page cache, which is effectively the set of all virtual memory pages currently in RAM.
The memory used by the disk cache is usually reported as unused, because the system can evict pages owned by the disk cache at will to free memory when applications request it. When nothing else needs the memory, the system may refer to data in the disk cache (when the desired data is indeed in the cache) and see large speedups.
In order to keep I/O operations fast, the system might opt to swap pages owned by one of your programs out to disk (on a UNIX system, for example, cron might be a good candidate to swap out) in order to make the disk cache larger. This means that the program will be slower to resume execution, but the system as a whole will be faster until then.
So, what happens when you have a very slow storage device (such as a USB thumb drive) connected? Your operating system usually tries to cache operations on that device a lot, since it’s very slow3. If you write a lot of data to it, the write will seem to complete much faster than is actually did, because the system is caching those writes in RAM and actually writing to the device as possible, which is rather slowly.
So what happens when a running program requests memory from the system and RAM is full of dirty pages waiting to be written to your slow device? The kernel will start evicting the dirty pages and writing them to your very slow device. Meanwhile, whatever program requested the memory is still waiting for it, unable to act. The act of requesting memory causes a page fault, preempting the program until the kernel can fulfill the request.
As a real-world case (the investigation of which prompted me to write this), when I’ve been transferring large amounts of data to a slow USB device and attempted to do other things, my system has come to a painful halt while the kernel has flushed dirty pages over USB when another application (such as Firefox) has requested more memory. It takes a long time for some memory to become available for Firefox, during which time it fails to accept any input, since the kernel is busy making space in main memory for it. By adding some swap on a faster device in the system, it can swap some other things out and keep everything running.
Now that we have established that it’s a good idea to have some amount of swap available, how much should you have? There’s no point in wasting disk space for swap you’re never going to use, but you want enough to be useful.
The old guideline for Linux was to allocate twice as much swap on disk as you have physical memory. On a modern system, following that rule can consume large amounts of disk space to no useful effect. While storage density has increased with time in a fashion similar to main memory capacity, the amount of space required for swap has greatly increased relative to the amount of data a typical user will need to store.
Taking my desktop workstation for example, it currently contains about 3 gigabytes of data which I consider irreplaceable, in a total of about 1.5 terabytes of total data. If I were to follow the old recommendations, the 16 GiB swap space would dwarf the data which I truly value. On a system with several terabytes of total storage, that’s not a lot of space, but many gigabytes of storage space seemingly going to waste doesn’t sit well with many users (myself included). There’s got to be a better approach, and indeed there is.
A good, simple rule of thumb for allocating swap space on a modern system is to have as much swap as you do physical memory. For typical operating conditions, that’s still going to be much more swap than you will ever need. However, that normally-unused space is not going to waste, because Linux uses your swap space to store the contents of main memory when the system suspends to disk (‘Hibernation’ in the parlance of most end-user systems).
Of course, some users don’t ever use the system’s suspend-to-disk functionality. For those situations, you’ll need to use your own discretion and strike a balance between wasted space and resistance to memory pressure. More swap will allow the system to absorb periods of high memory pressure more gracefully (and possibly avoid invoking the OOM killer in extreme situations), but will be mostly idle in the lulls.
For my part, I find a 2 GB swap partition to be acceptable on my previously-mentioned desktop workstation, yielding a total virtual memory capacity of roughly 10 GiB. 10 GiB is more memory than I anticipate ever needing for the system (barring the occasional buggy prototype piece of software, in which case I should fix it anyway), but the relatively small amount of disk-backed virtual memory can allow I/O-heavy workloads to push applications which aren’t in use out to disk to improve I/O performance.
You may be wondering now, “But I have a solid-state disk in my system! Won’t putting swap on that wear it out?”4 You’re right, but it’s probably not enough to be concerned about.
Newer SSD controllers are very intelligent about wear-leveling in order to extend the life of the memory, which is already pretty good (on the order of 100000 writes). On a 64 GiB device, that translates to an overall write endurance of 64 petabytes before the backing memory is running out-of-spec. If you write 64 GiB to it per day, it would require over 2000 years of operation to reach that point.
If you’re still paranoid about the lifetime of your storage device, you could always push swap onto a spinning disk, which sidesteps the write endurance problem at the cost of some performance.
As always, there are going to be cases where the ‘best practice’ (having swap) is not actually going to be best for you. However, I have tried to outline the facts here, to allow readers to make informed decisions.
In the overwhelming majority of situations, you should have some swap available, and if you’re not happy with exactly how the system handles it, Linux has some knobs for tweaking that behavior.5
Just to illustrate why understanding what’s going on and using personal discretion is important, a counterexample:
My netbook doesn’t need swap, and really shouldn’t have any. I only have one storage device in the machine, which is a very slow SSD with what is likely an extremely simple controller (that does little to no wear-leveling). Because the primary storage in that machine is so slow, I’m probably never going to see any reasonable performance improvement by putting a swap partition on it. The machine does have a couple SD card slots, however, and it might be reasonable to put swap on an SD card.
A midrange modern processor cycles 2 billion times per second (or more), and a typical consumer-grade hard drive will take around 7ms to seek on a good day, meaning you waste 1.4 million instruction cycles waiting for data from the disk. By comparison, something in main memory on the same processor can be accessed in about 60 nanoseconds, or 120 clock cycles. ↩︎
Oftentimes only a few megabytes per second. ↩︎
Flash memory, which backs most solid-state storage on the market right now, can only be written some amount of times before it wears out and stops working properly. ↩︎