Difference between revisions of "About Compression"

From eLinux.org
Jump to: navigation, search
(add to category: System Size)
(Spellchecl)
 
Line 12: Line 12:
 
=== Space impact ===
 
=== Space impact ===
  
Generally compression will cause that less space is needed. Typically a piece of data that is compressed becomes smaller. Note though that this is not always the case. If you compress something which is already compressed there is probably nothing to gain, and in most cases you'll loose a little bit, because the compressor itself adds info for decompression. Compressing audio or video files (like mp3, DivX or jpeg files) also will yield little space reduction as these data files are internally already highly compressed.
+
Generally compression will cause that less space is needed. Typically a piece of data that is compressed becomes smaller. Note though that this is not always the case. If you compress something which is already compressed there is probably nothing to gain, and in most cases you'll lose a little bit, because the compressor itself adds info for decompression. Compressing audio or video files (like mp3, DivX or jpeg files) also will yield little space reduction as these data files are internally already highly compressed.
  
 
The amount of compression also depends on the compression algorithm and the options that are applied.
 
The amount of compression also depends on the compression algorithm and the options that are applied.
Line 24: Line 24:
 
So whether or not compression affects your performance positive or negative depends on factors like read speed and decompression speed.
 
So whether or not compression affects your performance positive or negative depends on factors like read speed and decompression speed.
  
Of course it is not as simple as this. If reading is done through DMA and your system is CPU constrained compression will cause performance degradation. This is because reading using DMA is almost free as far as the CPU is concerned, and the decompression additionally loads the CPU (which was already the bottlenect).
+
Of course it is not as simple as this. If reading is done through DMA and your system is CPU constrained compression will cause performance degradation. This is because reading using DMA is almost free as far as the CPU is concerned, and the decompression additionally loads the CPU (which was already the bottleneck).
  
 
== Where compression may be applied ==
 
== Where compression may be applied ==

Latest revision as of 22:00, 27 October 2011

About Compression

Introduction

This page discussed compression in relation to boot and load time.
First we explain the effects of compression on the system in general.
Next we discuss where compression may be applied.
After that we will discuss the alternatives and go into detail whether they are worthwhile or not.

Effects of compression (general discussion)

Space impact

Generally compression will cause that less space is needed. Typically a piece of data that is compressed becomes smaller. Note though that this is not always the case. If you compress something which is already compressed there is probably nothing to gain, and in most cases you'll lose a little bit, because the compressor itself adds info for decompression. Compressing audio or video files (like mp3, DivX or jpeg files) also will yield little space reduction as these data files are internally already highly compressed.

The amount of compression also depends on the compression algorithm and the options that are applied.

Performance impact

Of course compression also has a performance impact. At first sight it may seem that the impact is negative, as additional time is needed to compress/decompress the data. However experiments may yield otherwise.
Consider the following situation. You have a 2 MB linux kernel. Compression will reduce that kernel to 1 MB. Now suppose you store the kernel in flash which has a read speed of 20 MB/s. In case of an uncompressed kernel, 100 ms will be needed to read the kernel from flash to RAM. However if the kernel is compressed, the actual I/O activity will only take 50 ms. So if decompressing 1MB takes less than 50 ms (decompression speed > 40 MB/s) you gain some time.

So whether or not compression affects your performance positive or negative depends on factors like read speed and decompression speed.

Of course it is not as simple as this. If reading is done through DMA and your system is CPU constrained compression will cause performance degradation. This is because reading using DMA is almost free as far as the CPU is concerned, and the decompression additionally loads the CPU (which was already the bottleneck).

Where compression may be applied

There are a number of places where compression can be applied in Linux:

  • If you use an initramfs in the kernel by default this initramfs is compressed.
  • The kernel itself often is also compressed. (hint: if your kernel name is vmlinuz or bzImage it is most likely a compressed kernel, if your kernel name is vmlinux it is probably not compressed; if you want to make sure: file is your friend here.
  • The filesystem itself can also apply compression. This is often the case with flash file systems. Notably SquashFS, CRAMFS and JFFS2 use filesystem compression.
  • Compression of executables, application data and the like. This is outside the scope of this page and not covered.

Initramfs compression

Initramfs in most cases is compressed. If you specify an initramfs file system while building the kernel, the initramfs image will be compressed and embedded into the kernel image. It is not required to compress the image, but there is no CONFIG option to disable this. The build scripts always perform the compression, and only if you tweak the script you can avoid the compression.

Another way to specify an initramfs is through the initrd= boot line parameter. In that case it is totally up to you whether or not you use a compressed or uncompressed cpio archive.

Now is initramfs compression useful? Well the first answer depends on load time versus decompress time as explained before. However, if you are using a compressed kernel and an embedded initramfs image, compression is useless as you compress the data twice. First as initramfs image and later a second time when the kernel is compressed. This is not giving additional benefit and only wastes CPU time.

Despite the arguments above though, there is one case where a compressed initramfs image in a compressed kernel is meaningful. That is if you are low on RAM. A compressed initramfs requires less RAM so it might well be that a compressed image fits whereas an uncompressed image does not fit. Note though that this only applies to RAM usage during boot. When the bootstrapping is completed and the kernel is started the space taken up by the initramfs image is released.

Kernel compression

For kernel compression the general arguments hold. However note that kernel decompression is done by the boot loader. This means that factors like read speed and decompression performance depend on the implementation of the boot loader and not on the implementation in the kernel. As these normally differ decompression speed here might differ from the decompression done in the kernel for initramfs.

Filesystem compression

The underlying filesystem can also provide compression. So you could in theory have a compressed initramfs in a compressed kernel which reside in a compressed filesystem. Not desirable at all. If your filesystem uses compression you probably do not want to compress kernel or initramfs.
Then again this is probably not likely. Most boot loaders seem to load the kernel directly from flash or from an uncompressed filesystem.

Concluding remarks

  • Take care to avoid nested compression
  • When deciding whether to apply compression or not, make sure to benchmark both alternatives