Changes/SwapOnZRAM

swap on zram

Summary

Swap is useful, except when it's slow. zram is a RAM drive that uses compression. Create a swap-on-zram during start-up. And no longer use swap partitions by default.

Owner

• Name: Chris Murphy
• Email: chrismurphy@fedoraproject.org

Current status

• Targeted release: Fedora 33
• Last updated: 2020-06-04
• FESCo issue: <will be assigned by the Wrangler>
• Tracker bug: <will be assigned by the Wrangler>
• Release notes tracker: <will be assigned by the Wrangler>

Detailed Description

zram Basic function

The zram† device, typically /dev/zram0, has a size set at create time during early boot, by zram-generator† per its configuration file. The memory used is not preallocated. It's dynamically allocated and deallocated, on demand. Due to compression, a full /dev/zram0 uses half as much memory as its size.

The /dev/zram0 behaves like any other block device. It can be formatted with a file system, or mkswap, which is the intention with this change proposal.

The system will use RAM normally up until it's full, and then start paging out to swap-on-zram, same as a conventional swap-on-drive. The zram driver starts to allocate memory at roughly 1/2 the rate of page outs, due to compression. But, there is no free lunch. This means swap-on-zram is not as effective at page eviction as swap-on-drive, the eviction rate is ~50% instead of 100%. But it is at least an order of magnitude faster than drive based swap.

zram has about 0.1% overhead or ~1MiB/1GiB. If the workload never touches swap, this overhead is the sole cost. In practice when not used at all, feature owner has experienced ~0.04% overhead.

Example: A system has 16 GiB RAM. The proposed defaults suggest the /dev/zram0 device will be 4 GiB. If the workload completely fills up swap with 4 GiB of anonymous pages, what's happened? The zramctl command will display the true compression ratio. If 2:1 is really obtained, it means 4GiB swap data is compressed to 2GiB. Therefore 2GiB is the actual RAM usage, and is also the net effective eviction. i.e. 4 GiB anonymous pages are evicted, but are then compressed and pinned into 2 GiB RAM, for a net memory savings of 2 GiB.

†
kernel.org zram.txt

Github zram-generator project

Overview of the Feature

Using swap is a good idea†, but no one likes it when it's slow. Anaconda and Fedora IoT have been using swap-on-zram by default for years. This builds on their prior effort.

There are three components to the change:

1. Install systemd rust-zram-generator† package. This does not enable swap-on-zram, it only makes the generator available.
   
2. Install a default zram-generator configuration. When present, swap-on-zram is set-up during startup.
   
3. Do not create swap partition/LV with default installations.

This proposal aims to apply all three, for all Fedora editions and spins, by default.

It further aims to apply the first two, for upgrades and custom installations.

It might be useful to only make the generator available (1), should an edition/spin wish to opt out, or as a fallback if applying the feature to upgrades fails to withstand scrutiny.

†
There is a tl;dr section at the top. Highly recommend reading the whole article. In defence of swap: common misconceptions

Default zram device configuration:

During startup, create a zram device /dev/zram0, with a size equal to 50% RAM, but capped† to 4 GiB, and with a higher than typical swap priority†.

These values seem reasonably conservative, and are based on prior work in Fedora. Anaconda sets swap-on-drive sized to 50% RAM in the no hibernation case, common outside x86. Fedora IoT's implementation also sets swap-on-zram size to 50% RAM.

†
RFE: should be able to set a cap on zram device size #10

RFE: should set priority #8

Default installer behavior

The installer is currently responsible for creating a swap-on-drive device. This will be dropped. The zram-generator + configuration file will trigger the setup and activation of swap-on-zram. This means hibernation isn't possible, even on systems that could support it.

Please see Supporting hibernation in Workstation edition for much more detailed information, including why it's increasingly likely hibernation isn't possible anyway, and a path to improving hibernation support.

Custom/Advance partitioning installer behavior

The user can add swap using Custom partitioning at install time. This is swap-on-drive. And the installer will also include the resume=UUID kernel parameter for this swap device. No change in behavior here.

Since swap-on-zram is still enabled by default, there will be two swaps: swap-on-zram, and swap-on-drive. The swap-on-zram will have higher priority, thus being favored over drive based swap. The kernel is smart enough to know it can't hibernate to a zram device, and will instead use drive based swap.

How can it be disabled?

Immediately:
swapoff /dev/zram0

Permanently:
rm /etc/systemd/zram-generator.conf

Feedback

You're enabling it on upgrades?

That's the current plan. As a technical matter, feature owner is confident this feature will improve the experience of all users regardless of configuration. As a non-technical matter, it's recognized that (a) hey pal, you're messing with my customizations, not cool! and (b) swap always stinks, I don't care if it has a 'Z' in the name! may need more convincing.

There are possible risks.

• Workloads that expect full use of memory, and depend on 100% page eviction. These may run slower if they really need full use of memory, but some memory is used for the zram device instead. Such workloads might favor zswap.

• Workloads with low compressible pages. In the worst case, this means unnecessary work merely moving pages around.

• Workloads with memory full, and hibernation. Hibernation is already stressful to memory-management subsystem and prone to bailing out in such cases. The swap-on-zram will be favored for evictions in the attempt to free memory to create the hibernation image. It could increase instances of hibernation entry failure. This isn't a crash, it just means the attempt doesn't succeed, and the system resumes operation instead of hibernating.

While possible, it's difficult to estimate their probability. But this is a significant consideration in the conservative default zram size. Users can easily increase zram size as needed for their use case, simply by editing /etc/systemd/zram-generator.conf and the change takes effect at next boot.

Why systemd zram-generator?

It's the most upstream implementation to date, is fast and lightweight. The zram-generator uses existing systemd infrastructure to setup the zram block device, format it as swap, and swapon - all during early boot. It's very similar in behavior to fstab-generator, gpt-auto-generator, and cryptsetup-generator†.

Converging on one implementation avoids user confusion. And while the alternatives are nice and work fine, a systemd generator is particularly well suited for this use case compared to a systemd service unit.†

Also, it's an reference implementation of a system generator written in Rust.

†
freedesktop.org About systemd generators.
devel@ Re: swap-on-zram by default Zbigniew Jędrzejewski-Szmek, systemd zram-generator author/maintainer

Why not a bigger zram device?

The main idea of being conservative is to address concerns about upgrades. It's possible some workloads will have less compressible data. Hence, not going with /dev/zram0 sized to 100% of RAM at this time. Even a /dev/zram0 of 200% RAM is not unreasonable *if* the compression ratio is at least 2:1. However, it's possible a system can get "stuck" in a kind of swap thrashing similar to conventional swap-on-drive, except it's CPU and memory bound, rather than IO bound. Feature owner thinks it's better to just oom, instead of getting overly aggressive with the zram device size.

Conversely it's possible to be too conservative with the size, and result in more instances of OOM kill. If applying the feature to upgrades is rejected, it's probably reasonable to increase the cap to ~8GiB. Of course more feedback and testing is needed, and it will be taken into consideration.

Note that the kernel zram doc says an excessively sized zram device does come with overhead. Users's can increase the size easily post-install, a capability they don't easily have with swap-on-drive. The goal for Fedora 33 is a default that's useful and safe for the vast majority of use cases.

Why not zswap?

Zswap† is a similar idea, but with a totally different implementation. It is swap specific, uses a RAM cache, and requires a conventional swap partition existing already. It might be true certain workloads are better suited for using zswap. But swap-on-zram depends only on volatile storage. This is simpler and it's more secure. Whereas zswap "spills over" into swap-on-drive and will leak user data if that swap device isn't encrypted. Some workloads may do better with zswap, and it's a valid future feature for a new generator, or possibly extend zram-generator to support it via the configuration file. Maybe the generator could favor zswap when swap-on-drive already exists; and fallback to swap-on-zram?

†
kernel.org zswap.txt

Benefit to Fedora

• significantly improves system responsiveness, especially when swap is under pressure;
• more secure, user data leaks into swap are on volatile media;
• without swap-on-drive, there's better utilization of a limited resource: benefit of swap without the drive space consumption;
• complements on-going resource control work, including earlyoom;
• further reduces the time to out-of-memory kill, when workloads exceed limits;
• improves performance for both "no swap" and "existing swap" setups;

Scope

• Proposal owners:
  • add zram-generator package to comps and kickstarts as appropriate
  • obsolete zram package (used by Fedora IoT)
  • means of per edition/spin configurations, if needed
  • coordinate a test day

• Other developers:
  • Anaconda are agreeable to deprecating their built-in implementation in favor of swap-on-zram
  • RFE's for zram-generator: users are not worse off if they don't happen. Open request for help, to make it possible. It's much appreciated.
    

RFE: should be able to set a cap on zram device size #10
RFE: should set priority #8

• Release engineering: #9495

• Policies and guidelines: N/A

• Trademark approval: N/A

Upgrade/compatibility impact

Add Supplements:fedora-release-common to zram-generator to pull it in on upgrades.

Existing systems without swap will have swap-on-zram enabled.

Existing systems with swap-on-drive, will also have swap-on-zram enabled (two swap devices), with higher priority for the zram device. Existing swap-on-drive will not be removed.

'zram' package contains zram-swap.service and associated bash scripts, and is currently used by Fedora IoT and ARM spins. It will be obsoleted to avoid conflicting/duplicative swap-on-zram implementations.

How To Test

Any hardware. Any version of Fedora.

1. dnf install zram-generator
2. cp /usr/share/doc/zram-generator/zram-generator.conf.example /etc/systemd/zram-generator.conf
3. Edit the configuration
4. Reboot
5. Check that swap is on a zram device: zramctl, swapon
6. Detailed check: journalctl -b -o short-monotonic | grep 'swap\|zram'
7. Check that priority is higher than existing swap if two or more are listed. ## (Enhancement is needed for this.)

Suggested configuration file values:
[zram0]
memory-limit = none
zram-fraction = 0.5

Feel free to run your usual workloads more aggressively or in parallel. Suspend-to-RAM and suspend-to-drive are expected to continue to work too (or at least hit all the same bugs as without zram being used).

Also, you can see the actual compression ratio achieved with the following command:
zramctl

Test Day

QA: SwapOnzram Test Day to discover edge cases, and tweak the default configuration if necessary to establish a good one-size-fits all approach.

User Experience

The user won't notice anything displeasing. If their usual workload causes them to dread swap thrashing, they'll be surprised that thrashing doesn't happen. The user might get curious if they don't find a swap entry in /etc/fstab. Or if they 'swapon' and see swap pointing to /dev/zram0 instead of a drive partition or LV.

Dependencies

N/A

Contingency Plan

• Contingency mechanism: Don't ship the generator = big hammer, but easy. Preferable to ship the generator, but only selectively ship configuration files = scalpel, pretty easy.
• Contingency deadline: Beta freeze
• Blocks release? No.
• Blocks product? No.

Documentation

Consider adding a hint in an /etc/fstab comment? There is no man page for this, and the documentation is also minimal, besides what's in this feature proposal. It's an open question how the user should get more information on how to configure and tweak it. But then, they don't have that for swap today either. There's just institutional knowledge.

Hence, a strong test day, with a lot of people and press coverage of the feature, might help spread the word for institutional knowledge changes coming.

Ideas welcome.

Release Notes

Pending feedback and test day.