After familiarizing yourself with swap, you might have a few questions. How much swap space? What do I set swappiness to? Do I even need swap anymore?

Note: This article is the third in a multipart series introducing the Linux swap. Part 1 is intended to familiarize the reader with the concept of swapping, why it exists, and what it’s used for. Part 2 highlighted basic analysis and management techniques for handling swap space. Part 3 appears below and discusses the current state of swap usage and present opinions about its implementation today.

Questions

Several questions are repeatedly posed in regards to swap:

• How much swap space do I need?
• What is the right settings for swappiness?
• With so much memory, do I even need swap?

I want to address each of these questions and hopefully provide a larger understanding of swap in the context of everyday use.

How much swap space do I need?

This is probably the most-asked question in regards to swap, and the answer isn’t entirely clear-cut, however, there are a couple of rules to be guided by.

In a strictly technical sense, the amount of swap you’ll need should be the total amount of virtual memory you’ll need minus system memory. In other words,

Total Virtual Memory - Memory = Swap

Traditionally, the recommendation was twice the amount of RAM. However, this was many years ago when 128MB (or less) might have been standard, resulting in a 256MB swap partition, totaling as much as 384MB of virtual memory. This pales in comparison to today, when 2GB is a standard amount on desktops. It is unlikely that a server with 16GB of RAM will ever need an additional 32GB of swap.

Desktop

Remember that virtual memory is a mapping of all available memory to a system, no matter the device type. Therefore, if you anticipate that you will never utilize more than 4GB of virtual memory and have 2GB of physical memory, 2GB might be a good amount for you. But better than 2GB of swap would be an additional 2GB of memory, allowing your heaviest usage to wholly fit inside RAM.

On my desktop at home (which currently runs Gnome + Compiz on Ubuntu 8.10 Intrepid Ibex), I started with 2GB of RAM and a 2GB swap partition. I’ve since upgraded the memory to 4GB and – as expected – seen swap usage drop to virtually none. The only situation in which my swap fills up, is when a process runs amok and grabs every last piece of memory available. Although I practically never reach into that extra memory, I was thankful it was there instead of having OOM Killer wreak havoc or forcing my system to a screeching halt.

Because the old “two times” model is bunk, some are recommending a new allocation guideline that goes something like this:

Laptop

Assuming similar usage patterns, you can follow the guidelines as a desktop, with one major exception. Because of the ability to hibernate – if you plan to use this functionality – allocate at least as much swap as you have of system memory. This is especially important during partitioning, because Linux cannot hibernate to a swap file.

Server

Generally speaking, servers tend to be loaded with memory for speed and efficiency, reducing the tendency to page to disk. Because they are often designed with a particular function and load in mind, swap may be of little need. In fact, if the server is adversely swapping, it is likely indicative of a more serious problem requiring more memory or application changes.

A consequence of this would be a small amount of swap (half, quarter, or less of system memory). When provisioning the server, you might instead decide that the workload and applications have been set in a very specific way (i.e. a VM server) and would prefer no swap at all. I wouldn’t recommend this, but it’s certainly possible from a technical and practical standpoint.

There are, of course, exceptions to these rules. If you anticipate gigantic workloads where you cannot reasonably fit the entire contents into memory (i.e. very large datasets on heavy-traffic, multi-user database systems), you may have no other choice than to allocate lots of swap space.

Red Hat’s own recommendations for RHEL5 are as follows:

What is the right settings for swappiness?

This is an entirely subjective setting, but there are a few strong opinions on what they feel is best. Because there is no way to anticipate the workload of an OS on installation, default swapiness setting falls somewhere in the middle. For example, on Ubuntu, the default value is 60.

Very high

One 2.6 kernel maintainer insists that a very aggressive setting is preferred (i.e. 100). His logic is that wasting a lot of memory on large, bloated applications is not an ideal use of that memory space; instead, force those pages out to disk, freeing up memory for useful items. The counter to this argument is that this produces the undesired behavior of delayed interactivity between the user and system when switching back to that application, because the operating system is busy reading the pages off disk back into memory. A further concern of this behavior is that cache is being given greater importance than applications – a useful scenario in servers and development workstations, but less so for general desktop use.

Very Low

For the average end-user, system responsiveness likely trumps cache performance, so there is a tendency for users that tweak swappiness towards the bottom. Though many have tried setting swappiness to 0, they found it unworkable; instead, many report a better experience with a setting of 10. A group of members have even proposed that this value be tuned lower by default in Ubuntu. Telling the kernel to swap as little as possible also has the effect of “as late as possible”. If memory cannot entirely cover the requirements of the next process, the kernel has no choice but to swap immediately, making loading the process slower. This speaks to the argument for a high setting if your system is typically memory-constrained.

With so much memory, do I even need swap?

Even if you decide that you have copious amounts of memory for the intended tasks, the inexpensiveness of disk space makes not having at least a small swap space (256MB) ludicrous. Additionally, the kernel can do strange things when there is no swap available, throwing out-of-memory errors and killing processes, making this decision a no-brainier. Disk is cheap, so allocate some space for swap, even if as a swap file rather than partition.

Final Thoughts

As with most things in life, the answer to how to configure swap is, “It depends”. Your specific situation largely dictates how much space and how quickly the kernel should page items out to disk. Much like a financial budget, you make a best estimation at first and then changes as the situation requires. Aside from a few specific instances, you can grow or shrink swap area as your needs change.

With the additional knowledge on what swap is used for and how to gauge its importance to your workload, you can start making a better-informed decision on how much you need and how to configure it. Take a look back at Part 2 and use those tools to measure how your system is responding to the hardware constraints and system settings. Even with all this in mind, there is a single overriding rule to swap space: more memory is better.