Optimizing system performance

When the root filesystem used on the device is a flash memory card or flash memory as will often be the case, the performance bottleneck will often be write access to the root file system. It is therefore beneficial for overall system performance to reduce unnecessary writing by using methods such as ramdisk file systems, tmpfs, and reducing system logging activity.

It is also possible to tweak the kernel and hardware devices to improve performance.

Reducing system logging activity
In a default distro install, system logging is often configured fully, suitable for a server or multi-user system. However, on a single-user system the constant writing the many system log files will result in reduced interactive system performance, and reducing logging activity will be beneficial for performance as well as the lifetime of the flash memory.

In a Debian Wheezy installation, the default system logger is rsyslog, and it configuration file is /etc/rsyslog.conf. In the rules section, the following logs are often enabled by default:

auth,authpriv.*                       /var/log/auth.log *.*;auth,authpriv.none        -/var/log/syslog cron.*                        /var/log/cron.log daemon.*                      -/var/log/daemon.log kern.*                        -/var/log/kern.log lpr.*                         -/var/log/lpr.log mail.*                        -/var/log/mail.log user.*                        -/var/log/user.log

There may also be rules for -/var/log/debug and -/var/log/messages, and |/dev/xconsole.

Note that kernel messages are logged in both kern.log and syslog, in addition to being available from the dmesg command from kernel memory. In a single user system, it is possible to disable most or all of these logs by placing a '#' character at the start of the corresponding lines. Logs can be re-enabled if it is necessary to debug a system problem.

Optimizing cache directories
Applications like browsers and window managers that use a disk cache may conform to the XDG Base Directory Specification standard. In that case, the environment variable XDG_CACHE_HOME defines the directory where cache files are stored. By setting this variable to a ramdisk location, it is possible to greatly speed-up the performance of certain browsers that otherwise stall with heavy writing to the disk-cache in flash memory. For example, in a Debian Wheezy installation with LXDE, there may be a configuration file called /etc/alternatives/x-session-manager. By adding the line export XDG_CACHE_HOME="/dev/shm/.cache"

to the start of this file, programs running in X conforming to the standard will be using the ramdisk in /dev/shm to store cache files. One browser popular on light-weight configurations that benefits form this is Midori. The default disk cache size in Midori is 100MB, you can lower this if you don't have much system memory available, but you can configure Midori to clean the cache on exit and thus give back the memory to the system in Tools->Clear Private Data.

Using tmpfs
To mount /tmp and /var/tmp (the directories for temporary files used by programs) on a ramdisk, you may be able to add the following lines to /etc/fstab:

tmpfs   /tmp        tmpfs    defaults    0 0 tmpfs   /var/tmp    tmpfs    defaults    0 0

Refer to more general documentation sources for your distro for more details.

Increasing available system memory at boot time
It is possible to increase available system memory at boot time by reducing reserved memory allocated to video subsystems. See Kernel arguments for more details.

Increasing the memory clock speed
Overclocking the memory in your device may result in stability problems. However, the default memory clock on your device may be set up for reduced power consumption with battery usage (on a tablet), and when battery lifetime is not of great concern (you leave power connected) you may be able to significantly increase the memory clock speed and as a result system performance.

There is a utility to read out the current DRAM settings. It is available at https://github.com/maxnet/a10-meminfo. The dram_clk parameter shows the current DRAM clock speed. The speed must be a multiple of 24 Mhz. The default speed for tablets is generally 360 MHz, boards may have a higher default speed. Although speeds as high as 480 MHz may appear stable at first, on heavy workloads especially using the Mali GPU errors may occur at higher memory clock settings. Most people find that 384 or 408 MHz run stable.

Because the memory clock has to be configured at the beginning of the boot process, it is not straightforward to change the memory clock speed. It involves recompiling u-boot-sunxi and re-flashing the bootloader on your SD-card. If something goes wrong it has the potential to corrupt the SD-card installation or make it unbootable. So only continue if you are sure what you are doing.

If you don't already have u-boot-sunxi installed, clone it using

git clone https:///github.com/linux-sunxi/u-boot-sunxi.git

Then go to the board/allwinner directory and select the configuration that most closely matches your device. For example. for a generic A10 1GB tablet you would pick a10_mid_1gb. Edit the file dram.c in the directory for your device and change the .clock parameter to the desired memory clock frequency. It must be a multiple of 24, usually in the range 360-480 MHz. Higher speeds may not be stable. Then run from the u-boot-sunxi base directory (replace a10_mid_1gb in the command below with the configuration that you selected):

make a10_mid_1gb

Now it's time to re-flash the bootloader on the SD-card. Run:

sudo dd if=./spl/sunxi-spl.bin of=/dev/mmcblkN bs=1024 seek=8 sudo dd if=./u-boot.bin of=/dev/mmcblkN bs=1024 seek=32

where /dev/mmcblkN is the device name of your SD-card (for example /dev/mmcblk0). After rebooting the new memory clock speed should have taken effect. Verify it by running a10-meminfo (see above).

Recompiling the kernel
By default the kernel may be configured with a large set of drivers to support a large number of configurations. It is possible to disable drivers that you don't need to reduce kernel size. The kernel may also be set up with a wide range of debugging options. Reducing debugging options will reduce the kernel size and improve performance a bit, although it may make debugging hardware or software problems harder. An advanced option for expert users that may be enabled is the 2G/2G memory split with high memory disabled, which theoretically could result in a measurable performance improvement.

Using a lower resolution graphics mode
The Allwinner chip's graphics framebuffer is stored in the same memory chips as system memory. As a result, the memory access required for screen refresh detracts from the overall system memory bandwidth and latency. By setting a lower resolution graphics mode, or a graphics mode with a lower pixel (color) depth, it is possible to measurably increase system performance.

When using the console, using a color depth of 16bpp may not be visibly different from 32bpp while improving performance. In X, performance will be improved but you will see banding on color gradients and acceleration options such as VE and Mali may not available in X and in console mode. G2D acceleration does support 16bpp.

You can change the graphics mode with Kernel arguments, changing script.bin, or after boot with utilities such as a10_display and a10disp (see Display).

The following tables show memory performance using the benchmark program tinymembench (https://github.com/ssvb/tinymembench‎) at different resolutions, pixel depths and memory clock settings on the same monitor.

NEON read bandwidth (MB/s):

Memclock (MHz):	 360	 384	 408 Display output disabled	1114	1159	1198 1280x720x16bpp		1075	1128	1195 1280x720x32bpp		1053	1078	1126 1920x1080x16bpp		1045	1089 1920x1080x32bpp		 902	 972	 991

NEON fill bandwidth (MB/s):

Memclock (MHz):	 360	 384	 408 Display output disabled 1176	1256	1330 1280x720x16bpp		1088	1171	1328 1280x720x32bpp		 972	1078	1090 1920x1080x16bpp		 934	1005 1920x1080x32bpp		 545	 548	 552

NEON copy bandwidth (MB/s) (unrolled copy prefetched (64 bytes step)):

Memclock (MHz):	 360	 384    408 Display output disabled	 650	 714	 723 1280x720x16bpp		 610	 650	 723 1280x720x32bpp		 554	 594	 636 1920x1080x16bpp		 538	 575 1920x1080x32bpp		 453	 517	 523

Memory latency in ns (block size 1MB, single random read):

Memclock (MHz):	 360	 384	 408 Display output disabled	 171	 165	 159 1280x720x16bpp		 173	 166	 159 1280x720x32bpp		 174	 168	 159 1920x1080x16bpp 	 175	 168 1920x1080x32bpp		 176	 169	 163

Memory latency in ns (block size 64MB, single random read):

Memclock (MHz):	 360	 384	 408 Display output disabled	 264	 254	 240 1280x720x16bpp		 265	 256	 240 1280x720x32bpp		 268	 256	 241 1920x1080x16bpp		 267	 257 1920x1080x32bpp		 270	 259	 247 As you can see, running at 1920x1080x32bpp really detracts from memory performance, especially write (fill) speed, which is also reflected in copy speed. Lowering the resolution and pixel depth, and increasing the memory clock improves system performance. Running headless (no display output) is fastest.

Changing the CPU frequency governor settings
The CPU frequency governor may not be setup for maximum performance. See Cpufreq.

Using an optimized Xorg driver
Your distro might be set up to only use the standard framebuffer driver for X. You can enable the sunxifb driver for X following the instructions in the section Binary drivers. You can enable G2D for 2D acceleration using the G2D engine.