Try this if you have low RAM, I lived with it for months when I had a broken DIMM and had to make do with 4 GB. The difference is incredible.
/etc/tmpfiles.d/zswap.conf
#Type Path Mode UID GID Age Argument
w /sys/module/zswap/parameters/zpool - - - - z3fold
w /sys/module/zswap/parameters/compressor - - - - lz4
w /sys/module/zswap/parameters/enabled - - - - 1
/etc/sysctl.d/00-swappiness.conf
vm.swappiness = 100
Depending on your workload you may increase swappiness to 200 with good results.
You need to set up some 8 GB of swap, it's mostly for accounting purposes and will barely get used so it can be anywhere. If you already have zram, disable zram, it's counter productive. Use the swapon command with no arguments to check if you have zram.
I think what you've configured is probably the way to go around in FHS-based distro - using config files. I am on Guix (btw), and I already have a swapfile partition configured when I set up the system configuration:
The size is allright, but the zswap config is crucial.
I can't figure out how to do it in guix with any certainty, perhaps you can enable it manually once to test and then research it better if you feel that it's worth it. Like this, as root:
In regular swap, when you're low on memory the kernel chooses some memory pages (low priority processes and least recently used) and writes them on disk on a file or partition. When the process that owns those pages need them back, the kernel goes fetch them one at a time. Since memory pages are 4KiB the speed on this depends on the 4k random access speed of your disk.
You can have more than one swap, and the order they are used depends on their priority, or on the order they were enabled in if you didn't specify any priority.
zram is a kind of swap space that, instead of writing to disk, compresses the pages and writes them back in RAM. You set an uncompressed size for it and if the pages don't compress well (usually encrypted on already-compressed data) then it will occupy the same amount of RAM. Since you can't tell in advance what the compressed size will be and there's no mechanism to stop it from filling up, you must be conservative on the size of zram. When the zram gets full all new pages will go to the next swap in priority order. This causes a problem where there's old data you don't care about taking RAM space in zram that cannot be reclaimed, and then your workload is going to regular swap which is slow.
zswap is a layer on top of swap, the technical name is "frontswap". For it to work you need to already have a swap configured. Before memory pages are written down on the swap file, they are compressed, and if the compression ratio is good enough the pages go to RAM instead of to the swap file. You set a compressed size for the zswap (by default, 20% of your total RAM) and when this limit gets full the least recently used pages are written to disk. The compression is so fast that you barely notice a hiccup while it's happening, it feels like you magically have 50% more RAM than before.
Answering your question, zswap is configured by kernel parameters, and now that you mention it it might work to put the parameters on the kernel cmdline instead of editing sysfs, this means configuring the boot loader and adding zswap.enabled=1 zswap.compressor=lz4 zswap.zpool=z3fold to the kernel cmdline.
I was troubled why zswap.enabled=1, zswap.max_pool_percent=50 and vm.swappiness=100 worked, but not zswap.compressor=lz4 and zswap.zpool=z3fold. Turns out that the kernel available in the substitute server has it disabled. I guess I'll use lzo and zsmalloc for now.
zswap has the same purpose than zram but a better design: it sends to disk the pages that cannot be compressed, and when it gets full it writes them back least recently used first. zram on the other hand keeps uncompressible pages around, and when it gets full all new pages go to disk and it makes the situation worse when you're low on RAM.
am i right that you can create zswap block devices and mount them like with zram? (eg. mounting /var/tmp)
zswap is a "front swap", it needs a backing swap to function that's crucial to the design. It automtically goes in front of all the swaps you have enabled.
You could probably put the backing swap on a loop device on a tmpfs, but I don't know how it will handle the loopback. It's a better idea to put it on disk. It can be a slow or write-limited disk, it will not get used much. You definitely should not use zram and zswap at the same time.
damn. when i do try out zswap i'll try and do this. but i do a lot of compiling and i prefer not to do all those reads and writes on my hard drive, so i will be sticking with zram for now.
thanks for the advice! it was hard for me to understand the differences at first when i first started looking at them so your comments are helpful.
Try this if you have low RAM, I lived with it for months when I had a broken DIMM and had to make do with 4 GB. The difference is incredible.
/etc/tmpfiles.d/zswap.conf/etc/sysctl.d/00-swappiness.confDepending on your workload you may increase swappiness to 200 with good results.
You need to set up some 8 GB of swap, it's mostly for accounting purposes and will barely get used so it can be anywhere. If you already have zram, disable zram, it's counter productive. Use the
swaponcommand with no arguments to check if you have zram.I think what you've configured is probably the way to go around in FHS-based distro - using config files. I am on Guix (btw), and I already have a swapfile partition configured when I set up the system configuration:
system.scm
Here's the size it was allotted:
The size is allright, but the zswap config is crucial.
I can't figure out how to do it in guix with any certainty, perhaps you can enable it manually once to test and then research it better if you feel that it's worth it. Like this, as root:
The
1 > enabledmust be written after the other two. If you don't have sysfs mounted at /sys, check with themountcommand where it is.A couple of mailing list threads that may help you do it "properly" in guix:
https://lists.gnu.org/archive/html/guix-devel/2023-02/msg00077.html https://lists.gnu.org/archive/html/guix-devel/2018-03/msg00332.html
Okay, so from my understanding, zram, zswap and normal swap are not one and the same, right? Isn't zswap a kernel argument, whereas zram a service?
Yes, more or less, they are closely related:
In regular swap, when you're low on memory the kernel chooses some memory pages (low priority processes and least recently used) and writes them on disk on a file or partition. When the process that owns those pages need them back, the kernel goes fetch them one at a time. Since memory pages are 4KiB the speed on this depends on the 4k random access speed of your disk.
You can have more than one swap, and the order they are used depends on their priority, or on the order they were enabled in if you didn't specify any priority.
zram is a kind of swap space that, instead of writing to disk, compresses the pages and writes them back in RAM. You set an uncompressed size for it and if the pages don't compress well (usually encrypted on already-compressed data) then it will occupy the same amount of RAM. Since you can't tell in advance what the compressed size will be and there's no mechanism to stop it from filling up, you must be conservative on the size of zram. When the zram gets full all new pages will go to the next swap in priority order. This causes a problem where there's old data you don't care about taking RAM space in zram that cannot be reclaimed, and then your workload is going to regular swap which is slow.
zswap is a layer on top of swap, the technical name is "frontswap". For it to work you need to already have a swap configured. Before memory pages are written down on the swap file, they are compressed, and if the compression ratio is good enough the pages go to RAM instead of to the swap file. You set a compressed size for the zswap (by default, 20% of your total RAM) and when this limit gets full the least recently used pages are written to disk. The compression is so fast that you barely notice a hiccup while it's happening, it feels like you magically have 50% more RAM than before.
Answering your question, zswap is configured by kernel parameters, and now that you mention it it might work to put the parameters on the kernel cmdline instead of editing sysfs, this means configuring the boot loader and adding
zswap.enabled=1 zswap.compressor=lz4 zswap.zpool=z3foldto the kernel cmdline.I was troubled why
zswap.enabled=1,zswap.max_pool_percent=50andvm.swappiness=100worked, but notzswap.compressor=lz4andzswap.zpool=z3fold. Turns out that the kernel available in the substitute server has it disabled. I guess I'll uselzoandzsmallocfor now.Huh, when I set it up
zsmallocwasn't finished and never deallocated. But it seems that today it's the right choice! Thanks.i find zram very useful (especially for fast temp editing of files) but that would only really work if you had enough RAM
zswap has the same purpose than zram but a better design: it sends to disk the pages that cannot be compressed, and when it gets full it writes them back least recently used first. zram on the other hand keeps uncompressible pages around, and when it gets full all new pages go to disk and it makes the situation worse when you're low on RAM.
woah thanks! i always thought zram was a better solution than zswap but i'll give zswap a try when i get time.
am i right that you can create zswap block devices and mount them like with zram? (eg. mounting /var/tmp)
zswap is a "front swap", it needs a backing swap to function that's crucial to the design. It automtically goes in front of all the swaps you have enabled.
You could probably put the backing swap on a loop device on a tmpfs, but I don't know how it will handle the loopback. It's a better idea to put it on disk. It can be a slow or write-limited disk, it will not get used much. You definitely should not use zram and zswap at the same time.
damn. when i do try out zswap i'll try and do this. but i do a lot of compiling and i prefer not to do all those reads and writes on my hard drive, so i will be sticking with zram for now.
thanks for the advice! it was hard for me to understand the differences at first when i first started looking at them so your comments are helpful.