svn commit: r250565 - head/etc

[Alex Kozlov]

On Mon, May 13, 2013 at 01:41:28AM +0400, Dmitry Morozovsky wrote:
> On Sun, 12 May 2013, Ed Schouten wrote:
> > 2013/5/12 Dmitry Morozovsky <marck at rinet.ru>:
> > > I'm afraid it could produce more harm than goodness on old hardware
> > > and/or other architectures like arm.
> > Any change we make at FreeBSD may or may not cause problems on old
> > hardware and/or other architectures like ARM. It's typically a case of
> > trial and error to see what happens.
> > In fact, I think that for embedded systems, using xz compression would
> > even be better. Many of those systems are often more storage space
> > constrained than CPU constrained (e.g. a 200 MHz wireless device with
> > only 8 MB of flash).
> > I think it's a pity the change has been reverted without bringing any
> > hard data to the table.
I've tested this change on i386/amd64 (Pentium T4400, the slowest I have),
arm (Raspberry Pi) and very old mips (mips 24k).
The difference in size between bzip2 and xz compressed logs around 20%.
The difference in compression speed around 100%. But I believe that it is
acceptable trade-of because logs rotation is an infrequent event, default
log size make the absolute value of compression time negligible and xz
decompresses several times faster than bzip2.

Brief test results (compression, 1 thread):
cpu       xz
T4400     1Mb/s
rpi       0.8Mb/s
mips 24k  0.07Mb/s (bzip2 0.16Mb/s)   

> I would pretty much like more statistics about the issue as well; 
> unfortunately, all I have handy are x86 hardware, and most of embedded-like 
> systems aer amd64 atoms...
> 
> I think some testing should be done on different platforms before making any 
> kind of decisions.
Oh well, let's just wait another 5 years.


-- 
Alex