Depend on use case. For me, its very crucial. Context, I'm working on ML and run several processing where usually consume memory up to 1TB at my workstation. Since my RAM only 196GB, this will trigger OOM killer, without Swapfile.
I use HHHL PCI SSD to host my swapfile to withstand with rapid read-write activities
Sounds interesting, could you please describe some details about "lvm/btrfs/zfs make that less of an issue"? Because I use btrfs almost everywhere, so I really wanna know more details about this, or could you please tell me where I can learn something about that?
On ZFS, use a zvol as a swap partition. Easy to create, destroy, resize, have redundant copies for robustness, all of the configuration options you have for a data volume but use for swap.
I've done this on both FreeBSD and Linux, never had a problem. However, IIRC there might have been some issues in the past on Linux.
I have a 2GiB swap on my 32 GiB worksation, and swappiness set to the lowest level.
Why ?
Because if I ever reached that situation, things will crawl to a halt and I might as well reboot.
Also, 2G is so cheap I'm just paying the cargo-cult entity to not think about it.
On my 64 GiB server, I have 4GiB, same reason.
Kernel has actually swapped out some stuff, probably ubuntu bloat stuff that's never used anyway.
Need? No. I havent used a swap partition since I first got a machine with 16GB RAM. I write text files for a living and its been plenty.
Every year or so a program (usually firefox) would spring a leak and lock up the machine for a minute or two then the OOM killer would kill it. This would have happened with or without swap; I think it's faster without?
So haven't bothered with partitions for a while. But a couple of years ago, the dists starting adding swap files by default. Ok, I thought, I won't fight it.
I do keep it reasonable, one or two gig. No problems so far. NVME is pretty darn fast, though htop says I don't really use it.
Get enough RAM for your use case, double if on one of those soldered Macs. This served me well, we had a 2013 MBP in great working condition (because I disregarded the salesbot advice and got 16gb) still works great but the recipient wanted a new M2 this year.
Same here. I've been using a desktop with 4GB RAM (considered low-end these days, many phones have that), but no swap.
Doing so for years. No big issues other than the occasional website (scripts, use of top-heavy web frameworks or whatever) causing runaway RAM usage. Local apps (including some older games, electronics design stuff) run fine.
Most sensible argument for using swap I've seen is a memory leak on server hardware, where having swap keeps the server at least running, buying the admin time to figure out what's going on.
On my PC, if an app leaks memory, I want to see free RAM going down to the point where something will crash - asap (no swap). Giving me a good sense which app (or website) was the culprit.
Not (with swap), the whole system getting slower & slower, but with a delayed fuse & much fuzzier clues about which app caused the problem.
Pick workflow + OS + apps such that working set fits RAM in the machine, and be happy. If needed, change OS / apps / workflow, or add more RAM.
Not: take app that needs 16 GB to run acceptable, run it on 4 GB machine, and use swap to somehow make it work while the user experience makes you cry.
So, imho: no need for swap on modern machines. Unless (maybe!) it concerns an internet facing server & uptime is critical.
Do you hibernate? Then you need swap, and I think it has to be a partition. I do because sleep is broken on this system (it doesn't reliably wake) so I disabled it in systemd to prevent the couple of accidental sleeps I did.
Is hibernation typically done through the swap partition?
I thought most OSs would just be using specialised sleep state files given that the hibernation image would need to be a serialised dump of all allocated memory whether resident on physical or paged out to the swap space. By extension, with swap disabled hibernation could still happen as long as there's enough physical memory free to handle writing to the sleep image on disk.
I'm a very beginner in my self learning journey. My understanding is, that hibernation is the same as 'suspend to disk'. If that is true, memory is cleared and put in a memory on disk (eg swap). From what I found on the Arch wiki, it works with a partition and file.
Thanks for the links. I guess it makes sense to use swap space in aiding suspend to disk if some of the allocated memory is already paged out to it. I wonder though if in the [2] link you posted if you can have a swap partition for the purposes of hibernation, while also disabling swapping of memory while your machine is in-use.
> makes sense to use swap space in aiding suspend to disk if some of the allocated memory is already paged out to it.
Maybe I can learn something here. From my understanding, the RAM image moves completely to the swap space, since the RAM memory is shut down (cleared). I thought, that's why it's calles 'suspend to disk'. During hibernation most components are in a state of being almost shut down. And that would be why you need a disk that can hold data non-volatile (eg not the RAM memory). Edit: From your comment it reads like the swap space is only for assisting the process, while from my understanding that's the place where the whole RAM image resides during sleep.
> a swap partition for the purposes of hibernation, while also disabling swapping of memory while your machine is in-use.
I thought the exact same. Avoid tear of the SSD and security implications, while having the option of hibernation (including the security implications) in case of need.
> Maybe I can learn something here. From my understanding, the RAM image moves completely to the swap space, since the RAM memory is shut down (cleared)
I think your understanding is broadly correct. My point was centred on the nuance that the primary purpose of swap space is for the OS to move less frequently used parts of virtual memory to disk to allow more recently/frequently used virtual memory to occupy the physical RAM.
So theoretically, if you wanted to implement hibernation on top of such a system, you wouldn't need the OS to have swapping enabled. (Although from your examples, practically, in some cases at least you do)
Without swap space, the OS would still use virtual memory addressing and allocation, but it would only be translatable to physical addresses in RAM, and exhaustion of RAM would lead to the OS out of memory killer to start killing processes to free RAM up.
But you could still write a hibernation program to take everything in memory and write it to disk and restore it later even if you're not using the OS swap capability to provide programs with the illusion of more memory to use than RAM allows.
Windows has HIBERFIL.SYS, which you can't even move off C:, in addition to the configurable swap files on every drive. Linux has whatever you configure but will refuse to hibernate without a large enough single swap although that doesn't have to be as large as the memory.
Now in theory the OS could not swap and still hibernate. Windows might do that if you disable the swap files. Linux might do that on minimum swappiness but you could also swapon what you need before hibernation and swapoff after boot/wake.
NO. Swap is evil.
The swap on HDD-based systems
kills performance.
with SSDs you can afford constant interruptions
at cost of faster wearout(especially TLC/QLC).
The gaslighting about swap being neccessary/good or
fast is strange - most machines run order of magnitude
faster without swap once you get a moderate load(even SSD benefit).
Some exotic workload need lots of RAM but demanding
this swapping be enabled by default is bad, its utility is justified by
weird ideas that "some shitware needs swap to work, because it just is so leave it
on and don't change it".
Thanks! And you're absolutely right, actually my biggest concern when debating the need for SWAP partition is whether there are some programs that depend on the SWAP partition to work.
I use some swap on servers because software leaks memory all the time and I’m cheap. Compressed swap can reduce a lot a cloud bill, and the performance impact of storing old unused memory pages on persistent storage is fine.
No, we don't. If a process begins swapping 9999 out of 10 000 it is because something has gone wrong and it is better to terminate that process than destabilizing the whole system by using the disk as ram.
Having a big enough swap partition to get a kernel core dump is handy. If you hibernate to swap, that needs to be big enough too. Sometimes swap partition is better than swap file, but a partition is harder to resize.
Personally, I find utility in having something approaching max(512mb, 2xRAM) swap available. If you've got an older system, 2xRAM is a good rule of thumb, but not for larger ram. IMHO, swap usage falls into two categories: slow leak and fast leak. In a slow leak, size doesn't really matter, just that it's being used, you can set your alert at 10% (or 75% or wherever) and you'll hit it eventually regardless of underlying size, in the mean time everything works fine, no big deal or urgency. In a fast leak, the system tends to get really slow really quick, until you hit the swap limit and then things usually get broken and fast; maybe if you get in early you can kill things before it spirals, but chances are you won't, and you'll just need to wait until you hit the max and then pick up the pieces; having a small swap means the time of max swapping is pretty limited.
TLDR: a small swap partition gives you a clear indication of a slow leak, but doesn't prolong agony during a fast leak.
Also note, measuring used memory is hard, but measuring used swap is easy. If your system typically has 0% swap used (which it should!), and starts swapping, you know something changed to use more memory (or the kernel swap algorithms changed)
42 comments
[ 3.3 ms ] story [ 90.7 ms ] threadI use HHHL PCI SSD to host my swapfile to withstand with rapid read-write activities
Depending on budget (and how expensive you consider your / development time), just choose between:
a) Use smaller ML models / tweak configuration such that it actually fits into 196 GB. That's already a ffing humongous amount of RAM! Or
b) Bite the bullet, and add RAM to fit the ML model(s) you're using.
Your workflow sounds like running Win95x on a 1 MB machine. Painful! (if it works at all). Just add the RAM, or slap DOS + Win 3.x back onto it.
With NUMA swap etc. also becomes a bit more complicated (see also https://blog.jcole.us/2010/09/28/mysql-swap-insanity-and-the... )
I think the main question here is to use a swap partition or a swap file.
The main problem of swap partitions is how inflexible they are, but lvm/btrfs/zfs make that less of an issue.
On Solaris a crashdump was written to swap when the system panicked. Not sure what Linux/BSD crashdump systems use nowadays.
Yea, that's true.
Just recreate the swap with the desired size. Extra easy with LVM.
I've done this on both FreeBSD and Linux, never had a problem. However, IIRC there might have been some issues in the past on Linux.
On my 64 GiB server, I have 4GiB, same reason.
Kernel has actually swapped out some stuff, probably ubuntu bloat stuff that's never used anyway.
Every year or so a program (usually firefox) would spring a leak and lock up the machine for a minute or two then the OOM killer would kill it. This would have happened with or without swap; I think it's faster without?
So haven't bothered with partitions for a while. But a couple of years ago, the dists starting adding swap files by default. Ok, I thought, I won't fight it.
I do keep it reasonable, one or two gig. No problems so far. NVME is pretty darn fast, though htop says I don't really use it.
Get enough RAM for your use case, double if on one of those soldered Macs. This served me well, we had a 2013 MBP in great working condition (because I disregarded the salesbot advice and got 16gb) still works great but the recipient wanted a new M2 this year.
Personaly went with the Framework so not worried.
I thought write text files may need swap partition more than others?
My text files are typically under 100k and a good editor will pull larger ones in as needed. 0% of them are larger than RAM, ~16 GB.
Doing so for years. No big issues other than the occasional website (scripts, use of top-heavy web frameworks or whatever) causing runaway RAM usage. Local apps (including some older games, electronics design stuff) run fine.
Most sensible argument for using swap I've seen is a memory leak on server hardware, where having swap keeps the server at least running, buying the admin time to figure out what's going on.
On my PC, if an app leaks memory, I want to see free RAM going down to the point where something will crash - asap (no swap). Giving me a good sense which app (or website) was the culprit.
Not (with swap), the whole system getting slower & slower, but with a delayed fuse & much fuzzier clues about which app caused the problem.
Pick workflow + OS + apps such that working set fits RAM in the machine, and be happy. If needed, change OS / apps / workflow, or add more RAM.
Not: take app that needs 16 GB to run acceptable, run it on 4 GB machine, and use swap to somehow make it work while the user experience makes you cry.
So, imho: no need for swap on modern machines. Unless (maybe!) it concerns an internet facing server & uptime is critical.
I thought most OSs would just be using specialised sleep state files given that the hibernation image would need to be a serialised dump of all allocated memory whether resident on physical or paged out to the swap space. By extension, with swap disabled hibernation could still happen as long as there's enough physical memory free to handle writing to the sleep image on disk.
Swap space can be used for [...] suspend-to-disk support. https://wiki.archlinux.org/title/Swap#Activation_by_systemd
In order to use hibernation, you must create a swap partition or file [...]. https://wiki.archlinux.org/title/Power_management/Suspend_an...
Maybe I can learn something here. From my understanding, the RAM image moves completely to the swap space, since the RAM memory is shut down (cleared). I thought, that's why it's calles 'suspend to disk'. During hibernation most components are in a state of being almost shut down. And that would be why you need a disk that can hold data non-volatile (eg not the RAM memory). Edit: From your comment it reads like the swap space is only for assisting the process, while from my understanding that's the place where the whole RAM image resides during sleep.
> a swap partition for the purposes of hibernation, while also disabling swapping of memory while your machine is in-use.
I thought the exact same. Avoid tear of the SSD and security implications, while having the option of hibernation (including the security implications) in case of need.
I think your understanding is broadly correct. My point was centred on the nuance that the primary purpose of swap space is for the OS to move less frequently used parts of virtual memory to disk to allow more recently/frequently used virtual memory to occupy the physical RAM.
So theoretically, if you wanted to implement hibernation on top of such a system, you wouldn't need the OS to have swapping enabled. (Although from your examples, practically, in some cases at least you do)
Without swap space, the OS would still use virtual memory addressing and allocation, but it would only be translatable to physical addresses in RAM, and exhaustion of RAM would lead to the OS out of memory killer to start killing processes to free RAM up.
But you could still write a hibernation program to take everything in memory and write it to disk and restore it later even if you're not using the OS swap capability to provide programs with the illusion of more memory to use than RAM allows.
Now in theory the OS could not swap and still hibernate. Windows might do that if you disable the swap files. Linux might do that on minimum swappiness but you could also swapon what you need before hibernation and swapoff after boot/wake.
Could you please be more specific?
Personally, I find utility in having something approaching max(512mb, 2xRAM) swap available. If you've got an older system, 2xRAM is a good rule of thumb, but not for larger ram. IMHO, swap usage falls into two categories: slow leak and fast leak. In a slow leak, size doesn't really matter, just that it's being used, you can set your alert at 10% (or 75% or wherever) and you'll hit it eventually regardless of underlying size, in the mean time everything works fine, no big deal or urgency. In a fast leak, the system tends to get really slow really quick, until you hit the swap limit and then things usually get broken and fast; maybe if you get in early you can kill things before it spirals, but chances are you won't, and you'll just need to wait until you hit the max and then pick up the pieces; having a small swap means the time of max swapping is pretty limited.
TLDR: a small swap partition gives you a clear indication of a slow leak, but doesn't prolong agony during a fast leak.
Also note, measuring used memory is hard, but measuring used swap is easy. If your system typically has 0% swap used (which it should!), and starts swapping, you know something changed to use more memory (or the kernel swap algorithms changed)