Reducing Boot Time in Embedded Linux Systems
Virtually every embedded system has some type of bootloader, and there are many bootloaders from which to choose. Some of the more popular include U-Boot for PowerPC, MIPS and ARM processors, and RedBoot, which is frequently found on ARM processors. Most popular bootloaders today contain far more functionality than actually is required for the task of initializing a system. Indeed, they have become valuable tools used by developers during initial board bring-up and system development.
Modern bootloaders are packed with features, such as Flash erase and program utilities, memory management utilities, network capabilities for loading images and for self-configuring (DHCP and BOOTP for example), drivers for PCI, IDE, USB and support for various partition types and filesystems. Some even have scripting language support useful for development, manufacturing (production test, image load and so on) and system upgrade.
These features make bootloaders an indispensable tool during product development. However, the size of bootloader images has significant impact on boot time. Bootloaders are stored in nonvolatile storage media, most commonly NOR Flash. However, embedded systems rarely execute code directly from Flash, mostly because it is far too slow. Read times of DRAM are orders of magnitude faster than read times from Flash.
The first job of the bootloader is to load itself into DRAM and continue execution from there. Consider the operating environment of early boot code. There is no stack, no C context (meaning all this code is written in the processor's native assembly language), and quite often, the processor's instruction and data caches are not yet enabled. This means the size of the bootloader image, which needs to be copied into RAM, will have a major impact on startup time.
The quickest path to performance improvement in your bootloader is to keep it small. Remove features that are unnecessary in a production environment. Some bootloaders, such as U-Boot, make it easy to do this. Its features are driven by a board-specific configuration file that contains directives to enable or disable features. Your requirements will ultimately rule, but prudent trimming of all but the most essential features will yield significant savings in boot time.
When you build a Linux kernel image, it is virtually always compressed as one of the final build steps. It is the responsibility of your bootloader (or a small bootstrap decompression loader that is appended to your kernel image) to decompress the kernel image and place it into system memory. One of the single largest easy gains you can achieve is to remove the decompression stage. Some architecture/bootloader combinations don't bother to enable caches, making decompression take much longer. It is not uncommon to find systems that take several seconds to perform the decompression and relocation stage. Using an uncompressed kernel can significantly reduce your system boot time.
Linux distributions use initrd/initramfs (hereinafter referred to as simply initramfs) primarily as a tool to enable a generic kernel to be used across a huge variety of system configurations. It is the job of the initramfs to provide the necessary device drivers to enable the devices that are required to complete system boot. Because embedded systems often are restricted to limited well-known configurations, they usually can eliminate initramfs with a corresponding reduction in boot time. Furthermore, removing support for initramfs in the kernel results in a smaller kernel (and, thus, faster boot.)
If you compile a kernel with a “default” configuration, it often contains a vast number of features your system may not need. You may be surprised to discover how many features are enabled by default that your embedded system does not need. Spend some quality time with your favorite kernel configuration editor (menuconfig, gconfig and so on), and go through each and every kernel configuration parameter. Evaluate whether your system requirements can do without it. Yes, it may take you the better part of a day (or even more if you add in some research time), but your savings in boot time reduction can be substantial. Some examples of features found in many default kernel configs include IPv6, RAID, support for many filesystems you may not need, extended partition support and many more. There also may be numerous device drivers compiled into the kernel for devices that are not present in your system. They are harmless, but each driver runs initialization code, including registration functions, and some spend lengthy milliseconds (or more) in device probe routines for devices that are not present—precious boot time can be spent probing for non-existent devices.
- High-Availability Storage with HA-LVM
- DNSMasq, the Pint-Sized Super Dæmon!
- Localhost DNS Cache
- Days Between Dates: the Counting
- Real-Time Rogue Wireless Access Point Detection with the Raspberry Pi
- You're the Boss with UBOS
- The Usability of GNOME
- Linux for Astronomers
- Multitenant Sites
- PostgreSQL, the NoSQL Database