Skip to main content

Analyzing cases for and against setting swap space on cloud instances

To swap or not to swap? That is the cloud question.
Image
Considerations for setting swap on cloud instances

Photo by Rakicevic Nenad from Pexels

If you want to start an argument with a Linux user, ask about swap memory. Some praise it as a cushion or as a safety net while others disparage it as a crutch and a destroyer of system performance. Born in the 1960s, swap memory has evolved over the years on Linux to serve two essential functions:

  • Emergency memory when a system consumes all of its RAM and still needs more
  • A parking lot for rarely-used memory pages that take up valuable system RAM

Many Linux distributions, including Red Hat, recommend swap memory for all systems. However, if you look at most cloud instances from various distributions, you find that swap memory is absent.

[ You might also like: Linux troubleshooting 101: System performance ]

Overview of swap

Swap exists on most systems as a partition on a disk. After partitioning, administrators format the partition with mkswap, enable it with swapon, and the kernel instantly sees available swap memory. Systems without an available partition can use a swap file, which is just a file on an existing filesystem that is formatted with mkswap and enabled.

Both methods work well, but putting a swap on a partition leads to better performance since you skip the overhead of a swap file on an existing partition.

When system RAM is scarce, Linux can store memory pages in the swap to avoid killing processes and crashing the system. Disks are much slower than system RAM, and this reduces system performance until system RAM frees up. If memory usage continues climbing to the point where system RAM and swap are fully exhausted, the Out of Memory (OOM) killer appears and begins killing processes until enough RAM becomes available.

The long-standing advice on swap size is double the size of your system's RAM. For example, allocate 2GBs of swap for systems with 1GB of system RAM. Although this ratio works well for smaller systems, it doesn't scale up for systems with hundreds of gigabytes of system RAM.

The case for swap on cloud

The rise of microservices (small, interconnected services that form a large application) led companies to deploy a larger number of smaller instances. Smaller instances come with less system RAM, which increases the risk of the dreaded OOM killer showing up to kill processes until the system has enough free memory.

Adding swap to these systems helps in two ways:

  • First, processes can briefly burst out of the system RAM into swap during periods of high load. Administrators learn about these events from their monitoring systems, and they have time to investigate the reason for the burst while it happens on the system. In the case of memory leaks, developers can inspect the process to understand what went wrong while it's still alive. This can also be a signal that administrators need larger instance types for their application as it grows.
  • Second, the Linux kernel keeps watch over rarely used memory pages and sends them out to swap to preserve precious system RAM. The sysctl setting vm.swappiness controls the kernel's desire to push memory pages to swap. This could help a cloud instance by keeping the most active pages in system RAM while pushing rarely used pages out to swap memory.

The case against swap on cloud

The "pets vs cattle" analogy comes up frequently when talking about cloud instances and administrators banish swap since they can automatically replace a misbehaving instance with a new one. Over time, their monitoring metrics show an increase in instances needing replacement due to OOM killer taking their application offline. The solution could be larger instances or further investigation of memory leaks under load. Systemd will also restart these processes quickly when possible.

Choosing a swap memory location on cloud instances requires more thought and planning. On physical servers, swap on a locally attached Non-Volatile Memory Express (NVMe) disk is fast enough, but what about a cloud instance with external storage mounted, such as AWS' Elastic Block Storage (EBS)? Performance on EBS varies depending on which type of EBS you choose and your neighbors on the system. Swap on a remote storage system could lead to poor instance performance when the system overflows its RAM. Administrators may opt to omit swap and replace these instances when they overflow their RAM instead of wrestling with poorly performing server handling requests.

Finally, many cloud instances become part of Kubernetes and OpenShift clusters, and they have a challenge when dealing with swap memory. There is a long-running GitHub issue about handling swap memory appropriately since it throws off the memory accounting for containers. Workarounds exist, but swap memory generally gets skipped on these systems.

Deploying cloud instances with swap

Should you decide that provisioning swap memory benefits your applications, cloud-init gives you the ability to provision swap memory on the first boot using the mounts module. Just add a few lines to your existing cloud-config user data:

swap:
    filename: /swapfile
    size: auto
    maxsize: 4294967296

This configuration tells cloud-init to create a swap file at /swapfile with an automatic size that is equal to or less than 4GBs. The formula for automatic swap sizing is in cloud-init's source code:

formulas = [
    # < 1G: swap = double memory
    (1 * GB, lambda x: x * 2),
    # < 2G: swap = 2G
    (2 * GB, lambda x: 2 * GB),
    # < 4G: swap = memory
    (4 * GB, lambda x: x),
    # < 16G: 4G
    (16 * GB, lambda x: 4 * GB),
    # < 64G: 1/2 M up to max
    (64 * GB, lambda x: x / 2),
]

The recommended swap sizing from Red Hat differs slightly:

  • 2GB or less of system RAM: 2x RAM
  • Over 2GB to 8GB: 1x RAM
  • Over 8GB to 64GB: 4 GB minimum
  • Over 64GB: 4GB minimum

You can manually specify the swap memory (in bytes) with the size: parameter
in the cloud-config:

swap:
    filename: /swapfile
    size: 2147483648  # 2 GiB

Provisioning swap on an existing cloud instance

For existing instances, a swap file is often the easiest method for enabling swap memory. In this example, place a 2GB swap file in /swapfile:


# BTRFS only #################################################################
# We must disable copy-on-write updates for swap files on btrfs file systems.
# The 'swapon' step fails if you skip these steps.
truncate -s 0 /swapfile
chattr +C /swapfile
# BTRFS only #################################################################

# A 2 GiB swap file.
dd if=/dev/zero of=/swapfile bs=1MiB count=2048

# Set the correct permissions on the swap file.
chmod 0600 /swapfile

# Format the swapfile.
mkswap /swapfile

# Enable the swapfile.
swapon /swapfile

# Add it to /etc/fstab to enable it after reboot.
echo "/swapfile none swap defaults 0 0" >> /etc/fstab

Check to see if your swap memory is ready:

$ cat /proc/swaps
Filename        Type    Size      Used    Priority
/swapfile       file    2097148   0       -2

[ Free online course: Red Hat Enterprise Linux technical overview. ] 

Wrap up

Swap memory provides two valuable benefits:

  • A safety cushion when system RAM usage increases to dangerous levels
  • A parking lot for rarely used memory pages that are taking valuable space in the system RAM 

That safety cushion comes with a performance penalty since memory paging to disks is incredibly slow relative to system memory.

Cloud deployments rarely see swap memory deployed, but swap still has benefits there. Choosing to deploy swap allows for a safety cushion when applications misbehave, but it could cause an application to perform slowly until the memory consumption issues get resolved. Skipping swap removes the cushion but ensures that a misbehaving application immediately stops. Cloud administrators must have a plan in place to handle these situations.

Luckily, cloud-init appears on the vast majority of cloud instances, and it allows for swap file creation with only a few lines of YAML. Swap memory is also easy to configure after the deployment using simple shell commands or scripts.

Whether you choose to deploy swap memory or to go without, ensure you have a plan when the worst happens. Monitor systems appropriately and keep a holistic view of the application that they serve.

Check out these related articles on Enable Sysadmin

Topics:   Linux   Linux administration   Cloud  
Author’s photo

Major Hayden

Major Hayden works on the Continuous Kernel Integration (CKI) project at Red Hat. In his spare time, he blogs at https://major.io and learns more about amateur radio. More about me

On Demand: Red Hat Summit 2021 Virtual Experience

Relive our April event with demos, keynotes, and technical sessions from
experts, all available on demand.

Related Content

OUR BEST CONTENT, DELIVERED TO YOUR INBOX