System Size

Revision as of 07:44, 21 October 2021 by Tim Bird (talk | contribs) (Using automation to hyper-specialize code)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search


Here are some links to information and projects related to Linux system size.

Technologies for decreasing system size

Kernel size reduction

Another wiki at has information about renewed efforts (as of August 2014) to reduce the kernel size.

Configuration Options

The Linux-tiny patchset

  • The Linux Tiny patch set is a collection of patches which can be used to make the Linux kernel consume less space. The long-term goal of the Linux-tiny project is to mainline these patches. Several patches have been mainlined over the last few years, and work continues in this area.


Andi Kleen submitted a set of patches (May 2014) to reduce the size of the Linux kernel networking stack. See the submission thread here:

Andi states that the patches support 3 use cases:

  • full-featured network stack (default Linux network stack)
  • client-only stack - with reduced features but still compatible with normal user-space programs, and suitable for some uses
  • minimal subset for deeply embedded, which would require specialized user-space software

In order to get full size reductions, it is best to use these patches with LTO. Doing so results in network stack that requires about 170K to run (versus 400K for the default stack).

Compiler options for reducing kernel size

An LWN article talks about three gcc options to shrink the kernel.

Shrinking the Kernel with GCC

The first option is -Os which is already in the tiny kernel patch.

Since version 3.4, gcc offered a -funit-at-a-time option. This apparently made gcc do a much better job of inlining and dead code removal, reducing the size of both text and data. It depended on another inlining patch. According to gcc's manual, this option no longer does anything.

The -fwhole-program --combine option set is equivalent to grouping all source files and making all variables static. These options are still supported by gcc, but not longer offered in BusyBox configuration options. What happened?

Another option, -mregparm=3, seems to be x86 specific, it instructs the compiler to use registers for the first three function arguments. by John Rigby

See [1] for all available optimization switches. See Compiler_Optimization for more details on effects of optimization options.

Section garbage collection patchset

These patches can shrink kernel size by ~10% by improving dead code/data elimination at link time. They are being pushed to mainline. Due to a linker bug, their acceptance depends on a newer, fixed linker (will be in binutils-2.21). Good news are that the bug affects only certain architectures (parisc), so the patches are usable even with "old" linker.

Runtime size of kernel

Often, the focus of memory size reduction for the kernel is on the size of the statically compiled image for the kernel. However, the kernel also allocates memory dynamically when it runs. On loading, the kernel creates several tables for things like network and file system structures.

Here is a table showing different kernel hash tables, and their approximate size for a 2.6 kernel. (Table taken from page 25 of )

Hash Table memory < 512MiB RAM memory >=512MiB RAM
32b/64b 32b/64b
TCP established 96k/192k 384k/768k
TCP bind 64k/128k 256k/512k
IP route cache 128k/256k 512k/1M
Inode-cache 64k/128k 64k/128k
Dentry cache 32k/64k 32k/64k
Total 384k/768k 1248k/2496k
kernel stack size

There used to be a configuration option for reducing the size of the kernel stack for each process to 4K. By default (as of 2011), the default kernel stack size is 8K. If you have a lot of processes, then using 4K stacks can reduce the kernel stack usage.

Some notes about this are at: Kernel Small Stacks


In 2012, Tim Bird studied a few different techniques related to automatic size reduction and whole-system optimization. Specifically, he studied the following items:

  • link-time optimization of the kernel
  • syscall elimination
  • global constraints
  • kernel stack reduction

Tim also found some very interesting academic research on link-time re-writing and cold-code compression. Tim's work was presented at LinuxCon Japan in May, 2013.

A draft outline and completed slides for the talk are at System Size Auto-Reduction

Compressed printk messages

The open project proposal Compressed printk messages evaluated this technique in 2014. The results can be found on the Compressed printk messages - Results page.

Reduction Ideas and recent work

A group of developers is (as of 2014) doing continued size reduction work on the Linux kernel. A page has been set up to categorize recent work and ideas for future kernel size reductions. The page is: Kernel Size Reduction Work

File system compression

For read-only data, it is useful to utilize a compressed file system. The following are used heavily in embedded systems:

  • Cramfs and SquashFS, for block storage.
  • JFFS2 and its successor UBIFS, for flash (MTD) storage.

Note that Cramfs and Squashfs, due to their "write-only-once" nature, can also be used on MTD storage.

See the File Systems page for more information.

Shrinking your application

Compiler options for program size

You can use "gcc -Os" to optimize for size.

Stripping your program

You can use the 'strip' command to eliminate unneeded symbols from your application. The 'strip' command should be included with your toolchain, and may be architecture-specific. (I.e. you may need to run it with a toolchain prefix, like "arm-linux-strip")

Note that this makes debugging your application more difficult, because the debug symbols are no longer available.

By default, strip just removes debug symbols. You can remove everything but the essential symbols used for dynamic linking. To get the highest savings, use "strip --strip-unneeded <app>"

This can save a lot of space, especially if debug symbols were included in the build.

$ gcc -g hello.c -o hello
$ ls -l hello
-rwxrwxr-x 1 tbird tbird 6143 2009-02-10 09:43 hello
$ strip hello
$ ls -l hello
-rwxrwxr-x 1 tbird tbird 3228 2009-02-10 09:43 hello
$ strip --strip-unneeded hello
$ ls -l hello
-rwxrwxr-x 1 tbird tbird 3228 2009-02-10 09:43 hello

Now, compiles without debug symbols to start with:

$ gcc hello.c -o hello
$ ls -l hello
-rwxrwxr-x 1 tbird tbird 4903 2009-02-10 09:45 hello
$ strip hello
$ ls -l hello
-rwxrwxr-x 1 tbird tbird 3228 2009-02-10 09:45 hello

You can strip both executables as well as shared libraries.

There is a "super-strip" utility, which removes additional material from an ELF executable program (which 'strip' usually misses). It is available at: This program appears to be obsolete now. I couldn't get it to compile on Fedora 8.

Some information about stripping individual sections by hand, using the -R command is available at:

Hand-optimizing programs, for size

If you are very intent on creating small binaries, you can use some techniques to manually create the smallest Linux executables possible.

See A Whirlwind Tutorial on Creating Really Teensy ELF Executables for Linux

Library savings

Use of a smaller libc

Glibc is the default C library used for Linux systems. Glibc is about 2 meg. in size. Other C libraries are also available for Linux, and they offer varying degrees of compatibility and size savings. In general, uClibc is considered a very good alternative to glibc, for systems where size is an issue.

  • uClibc - small footprint but complete C library
  • dietlibc - another library to produce very small statically compiled executables.
  • klibc - very small library for use in init ram filesystems
  • eglibc - a version of glibc designed for embedded systems. Reduced footprint is one of the design goals.
  • musl libc - a lightweight, fast, simple, and standards-compliant C library
  • olibc - another C library optimized for size and performance, derived from Android bionic libc
  • Subset Libc Specification - CELF once considered the possibility of creating a subset libc specification. Some companies have also examined the possibility of modularizing glibc, so that parts of it can be made configurable. Preliminary research indicates that this could be a very difficult thing, since glibc has very messy function interdependencies.

Static Linking

If your set of applications is small, sometimes it makes more sense to statically link your applications than to use shared libraries. Shared libraries by default include all symbols (functions and data structures) for the features a library provides. However, when you static link a program to a library, only the symbols that are actually referenced are linked in and included in the program.

Library reduction

It is possible to reduce the size of shared libraries, by eliminating unused symbols.

MontaVista released a tool for library optimization. This tool scans the entire file system, and can rebuild the shared libraries for the system, including only the symbols needed for the set of applications in the indicated file system.

Care needs to be taken with this approach, since it may make it difficult to use add-on programs or or do in-field upgrades (since symbols required by the new software may not be present in the optimized libraries). But for some fixed-function devices, this can reduce your library footprint dramatically.


Deferred Library Loading

It is possible to reduce the RAM runtime footprint for a product, by lazily loading shared libraries, and by breaking up library dependencies. Panasonic did some research into a process called Deferred Library Loading, which they presented at ELC 2007.

See the Deferred Dynamic Loading (pdf) presentation.


You can save RAM memory by using some text or data directly from flash.

Kernel XIP

By executing the kernel in-place from flash, it is possible to save RAM space.

Application XIP

By executing applications in-place from flash, it is possible to save RAM space.

Data Read In Place (DRIP)

This is a technique for keeping data in flash, until it is written to, and then making a RAM page for it.

Size measurement tools and techniques

Kernel size measurement data

  • Bloatwatch - a kernel size regression analysis tool.
    • Bloatwatch provides a great amount of detail, and the ability to compare the size of kernel versions over time.

How to measure the kernel image size

  • to see the size of the major kernel sections (code and data):
size vmlinux */built-in.o
[tbird@crest ebony]$ size vmlinux */built-in.o
   text    data     bss     dec     hex filename
2921377  369712  132996 3424085  343f55 vmlinux
 764472   35692   22768  822932   c8e94 drivers/built-in.o
 918344   22364   36824  977532   eea7c fs/built-in.o
  18260    1868    1604   21732    54e4 init/built-in.o
  39960     864     224   41048    a058 ipc/built-in.o
 257292   14656   34516  306464   4ad20 kernel/built-in.o
  34728     156    2280   37164    912c lib/built-in.o
 182312    2704     736  185752   2d598 mm/built-in.o
 620864   20820   26676  668360   a32c8 net/built-in.o
   1912       0       0    1912     778 security/built-in.o
    133       0       0     133      85 usr/built-in.o
  • to see the size of the largest kernel symbols:
    • nm --size -r vmlinux
[tbird@crest ebony]$ nm --size -r vmlinux | head -10
00008000 b read_buffers
00004000 b __log_buf
00003100 B ide_hwifs
000024f8 T jffs2_garbage_collect_pass
00002418 T journal_commit_transaction
00002400 b futex_queues
000021a8 t jedec_probe_chip
00002000 b write_buf
00002000 D init_thread_union
00001e6c t tcp_ack

How to measure the memory usage at runtime

See Runtime Memory Measurement for a description of ways to measure runtime memory usage in Linux.

Also, see Accurate Memory Measurement for a description of techniques (and patches) which can be used to measure the runtime memory more accurately.

Linux size increase from 2.4 to 2.6

Linux increased in size by between 10% and 30% from version 2.4 to 2.6. This incremental growth in kernel size has been a big concern by forum members.

Please see the Szwg Linux 26Data page for supporting data.

GCC Code-Size Benchmarking

CSiBE is a code size benchmark for the GCC compiler. The primary purpose of CSiBE is to monitor the size of the code generated by GCC. In addition, compilation time and code performance measurements are also provided.


Case Studies

"Motorola reduction of system size (presumably for cell phones) using 2.4 Linux"

    • MotSizeReduction.ppt - this is a placeholder for this Powerpoint as it was too big to upload to the wiki. (why is this even here? TRB)


Linux on MicroControllers (M3 in this case)

Using automation to hyper-specialize code

Here's an interesting story about a group that wrote a special tool to hyper-specialize code for a game (for the "demoscene" environment, where every byte counts).

Reduced-size distribution efforts

Here are some projects aimed at building small-sized systems:

meta-tiny is my experimental layer where I'm looking at what we can
build with our existing sources and infrastructure. I've found that we
can cut the image size to about 10% of core-image-minimal without
changes to source code, but dropping a lot of functionality. We can get
to something like 20% while still maintaining ipv4 networking.

Other Tidbits on System Size

Memory leak detection for the kernel

Catalin Marinas of ARM has been recently (as of 2.6.17?) been posting a memory leak detector for the Linux kernel. It may get mainlined in the future. Here's a link to the LKML discussions around it:

How System Size may affect performance

It has long been theorized that reducing system size could provide a performance benefit because it could reduce cache misses. There does not appear to be hard data to support this theory on Linux, but this has been discussed on the kernel mailing list.

See this post by Linus Torvalds

Stripping down the filesystem of a desktop distribution

Here is a good document with tips on how to strip out unneeded files from a desktop distribution. The example distribution used here is Linux From Scratch, but the tips should work with many distributions.

Extremely-minimal systems

This section lists various efforts to produce the absolute smallest system possible.