Nat!'s Web Journal

This is not a very exiting day or post :)

bash-2.05a$ time ./gawk.sh 

real    0m16.849s
user    0m14.650s
sys     0m1.940s

Typisch Telekom. Drei Anrufe, drei verschiedene Antworten. Ob die schon Eliza dort haben ?

Ich poste das hier "vonne Firma" :)

I suspect that I can improve realloc some, as the memcpy routine is never inlined apparently on PPC, by wrtiting my own memcpy. Looked around the net and already found an Altivec memcpy. I wonder how much memory one needs to move, so that it becomes a win, as context switches for non-Altivec code are bound to become much slower. (Sames goes for using doubles in memcpy).

Looks like a Linux problem so far ARGH! (Can't be my mistake now, can it :))

I am none too happy with the UI of X-Code, so I mailed a megabyte worth full of suggestions to Apple on the X Code UI. Ok Ok, not 1 MB full of text, only four pages with lots of pictures.

I think I am not breaking any NDA here, by presenting a mock up on how i think X-Code should look like, as there are screenshots of X-Code available at Apple.

It will be interesting to see if there is any feedback at all. Best feedback would be of course to see the ideas appear in the final release. One can always hope. :)

 PID COMMAND      %CPU   TIME    FAULTS   PAGEINS  COW_FAULTS MSGS_SENT  MSGS_RCVD  BSDSYSCALL MACHSYSCALL CSWITCH
 1289 Mulle        13.7%  0:04.23  3998     0        0          387        387        13245      387         1564
 1293 Apple         7.1%  0:04.92  4244     0        0          1256       1256       13245      1256        1784

As you can see at the end of the run, the Mullocator actually allocated less pages (FAULTS) and did much less Mach Calls, which is good. But speedwise I am still stuck around 60%. I suspect I will not convince the Darwinists that zero filling pages on a consumer machine is a waste of effort. :)

COW_FAULT (mooh) is a copy on write fault by the way, just looked it up.

http://www.macintouch.com/o98security.html

Basically it goes like this. Program TextEdit.app is used to write some text. That text ends up in the back of a memory page. When the program ends, that page is reused in another program lets say LISP or some such program that saves out complete memory pages (dumps an image). Now if that memory page is only partially overwritten the data edited in TextEdit.app might very well be appended at the end of that image file and if you then send this image to someone else, you may have sent some data out you rather wouldn't have liked to.

This Office effect is dealt on other systems deal by ignoring it (is this true ?), but Mac OS X is conscientious and zeroes out the memory page.

In most not so long running programs the zero filling is absolutely noticable. It's effects tend to peter off, if a program settles into a malloc free pattern with a stable amount of memory used.

So there isn't much I can do about the Safari.app benchmark.

Now this ain't a cheap operation, so I figured I start using that when 4 MB worth of free memory have accumulated,

Here's a post from Jim Magee (IMO Apple's best guy) about msync from the darvin-developers mailing list:

for( i = 0; i < 10000; i++) 
   malloc( 8);

This conflicts with my tendency to procrastinate and to then take on something totally different and forget the old crap. :)

How's that ? Pretty easy requirements! Please compile and run the following code on a system that is close to doing nothing, turn of Mail, turn off iTunes and what ever else might be sucking up CPU time. Compile this with cc -o page-clean-test page-clean-test.m -framework Foundation.

Please send the output result to nat@mulle-kybernetik.com (Harvest be damn SPAM bot!) Please write me the kind of system you ran the test on. If you didn't use Mac OS X 10.2, please tell me also.

Here comes the source...

• Copied the DevDocumentation.pkg down locally from the Packages folder.
• Show Package Contents on it, moved down to the Info.plist and edited it with Property List Editor
• Set IFPkgFlagRootVolumeOnly to No
• Created a /tmp folder on the Volume I wanted to install to.