Does ZRAM or ZSWAP makes sense in WSL2?

[crramirez]

Windows has its own RAM compression system, and WSL2 ultimately runs inside Windows.

Does the Windows RAM manager compresses the Hyper-V RAM as well?

[crramirez]

Anyone?

[4-FLOSS-Free-Libre-Open-Source-Software]

ZRAM works very well inside WSL2. In the meantime, you could use the zRAM enabled wsl2 kernel https://github.com/Locietta/xanmod-kernel-WSL2?tab=readme-ov-file#usage

[ksnishat]

I could not set zram after installing this custom kernel. Can you please provide a small set of instructions?

[4-FLOSS-Free-Libre-Open-Source-Software]

Zram confirmed working

[mangkoran]

@4-FLOSS-Free-Libre-Open-Source-Software May we know how to enable/add zram in WSL?

[Artoria2e5]

A kernel that has zram or zswap included is not all it takes to get it! You gotta explicitly enable it.

For how to get zram going, https://wiki.archlinux.org/title/Zram should be it. It's a bunch of commands to set up though.

Aaaand that's why I recommend just using zswap. Add kernelCommandLine="zswap.enabled=1 zswap.compressor=zstd zswap.zpool=zsmalloc" to your wslcondig under [wsl2] and it should work. I'm going to test this tomorrow because Haskell Language Server keeps crashing the VM by blowing up the RAM.

[4-FLOSS-Free-Libre-Open-Source-Software]

I could not set zram after installing this custom kernel. Can you please provide a small set of instructions?

The configuration files is the wrong place.
No idea what you looking for.
No ZRAM options exists there.

@4-FLOSS-Free-Libre-Open-Source-Software May we know how to enable/add zram in WSL?

Locietta/xanmod-kernel-WSL2#142 (comment)

Aaaand that's why I recommend just using zswap. Add kernelCommandLine="zswap.enabled=1 zswap.compressor=zstd zswap.zpool=zsmalloc" to your wslcondig under [wsl2] and it should work. I'm going to test this tomorrow

Yes, zwap works in xanmod WSL2 Kernel as described. Just add in .wslconf kernelCommandLine="zswap.enabled=1"

[Artoria2e5]

@mangkoran Alright as promised, guide to getting zswap turned on (zram involves per-distro configuration (need to set up startup tasks!) and might cause inversion BS with WSL2's disk swap file)

1. get xanmod-kernel-WSL2 using the scoop command
2. code ~/.wslconfig to add

   [wsl2]
   kernelCommandLine="zswap.enabled=1 zswap.shrinker_enabled=1 zswap.compressor=zstd zswap.zpool=zsmalloc vm.swappiness=133"
   [experimental]
   autoMemoryReclaim=DropCache

3. wsl --shutdown
4. open your favorite wsl instance, I don't care
5. Run grep -r . /sys/module/zswap/parameters/ and check that you have /sys/module/zswap/parameters/enabled:Y
6. Do something that eats up your memory, e.g. open haskell language server
7. Run grep -r . /sys/kernel/debug/zswap/ to see how much zswap is saving you

Also useful links are:

• https://wiki.archlinux.org/title/Zswap, for general linux stuff about zswap
• https://www.whitewaterfoundry.com/blog/2025/3/7/get-more-from-less-how-zswap-optimizes-memory-in-wsl2, which is where the gnarly-looking vm.swappiness=133 came from
  • there's also a nice script

Why high swappiness?

Swappiness determines whether the kernel prioritizes emptying the file pages (by writing to disk, usually) or emptying the anonymous pages (by moving them to swap) when memory is short. As a result the ideal swappiness is (100 × (IO cost of emptying file page) / (IO cost of emptying anonymous page)).

The IO cost of emptying a file page comes two things:

• Potentially needing to write the disk if it represents a pending write, one write per page-emptying max.
• Potentially needing to read the disk once again if the file is needed later, one read per page-emptying cycle max.

The IO cost of emptying an anonymous page also comes from two things:

• Definitely needing to write the page to disk.
• Potentially needing to read the page from disk if it's needed later, one read per page-emptying cycle max.

The kernel default swappiness is 60, which is very reasonable assuming the same IO device is used for file and swap. After all it's two potential costs against one definitive cost plus one potential cost.

However, zswap and zram lowers the effective cost of reading and writing to swap. Zswap does it by acting as a LRU cache in front of swap. Zram does it by acting as the swap. So our cost for emptying an anonymous page becomes:

• Needing to store the page somewhere, which can be an area of RAM or the disk
• Potentially needing to read the page from an area of RAM or the disk

This creates a situation where triggering the swap system may be cheaper than file IO even without considering filesystem overheads. It is totally possible for a page to be instead put somewhere in the RAM, which for most intents and purposes have near zero IO cost compared to disk.

On top of that we have filesystem and other overheads for file IO. In the case of WSL, well we all know about the overhead of accessing Windows drives via "drvfs". So what does decreased swap IO cost and increased file IO cost mean? That's right, higher swappiness is only logical.

Example performance values

Workload: VS Code and Haskell Language Server

# ./zswap_stats
Zswap Statistics:
-----------------
Stored Pages: 502402
Pool Total Size: 533MB
Uncompressed RAM: 2.0GB
RAM Saved: 1.4GB
# free -h
               total        used        free      shared  buff/cache   available
Mem:           3.8Gi       3.6Gi       171Mi       1.1Mi       271Mi       226Mi
Swap:          4.0Gi       2.6Gi       1.4Gi
# grep -r . /sys/kernel/debug/zswap/
/sys/kernel/debug/zswap/stored_pages:502170
/sys/kernel/debug/zswap/pool_total_size:558354432
/sys/kernel/debug/zswap/written_back_pages:154615
/sys/kernel/debug/zswap/reject_compress_poor:0
/sys/kernel/debug/zswap/reject_compress_fail:505
/sys/kernel/debug/zswap/reject_kmemcache_fail:0
/sys/kernel/debug/zswap/reject_alloc_fail:0
/sys/kernel/debug/zswap/reject_reclaim_fail:143
/sys/kernel/debug/zswap/pool_limit_hit:0
# grep -r . /sys/module/zswap/parameters/
/sys/module/zswap/parameters/enabled:Y
/sys/module/zswap/parameters/shrinker_enabled:Y
/sys/module/zswap/parameters/max_pool_percent:20
/sys/module/zswap/parameters/compressor:zstd
/sys/module/zswap/parameters/zpool:zsmalloc

[4-FLOSS-Free-Libre-Open-Source-Software]

Not having zswap.compressor=zstd currently only lz4 supported.

• [Kenrel Config] enable more features Locietta/xanmod-kernel-WSL2#94 (comment)

[Artoria2e5]

Not having zswap.compressor=zstd currently only lz4 supported.

* [[Kenrel Config] enable more features Locietta/xanmod-kernel-WSL2#94 (comment)](https://github.com/Locietta/xanmod-kernel-WSL2/pull/94#discussion_r1864556770)

No, CONFIG_KERNEL_ZSTD=y is present, which allows zstd to be used when you ask for it.

[crramirez]

There is a new version of the script that compiles the official Microsoft Kernel adding zswap. It was updated for the kernel 6.x case with external vhdx modules disk.

https://github.com/dhanar10/wsl2-kernel-zswap