_Arthur (Arthur_@sympatico.ca) on 5/18/06 wrote:
---------------------------
>Jasjeet Sekhon tested his statistical package under Linux and MacOSX on the same
>hardware, and concluded that MacOSX hybrid kernel was poorly implemented, (re-spouting Johan DeGelas allegations).
>
>http://sekhon.berkeley.edu/macosx/intel.html
>
>Ridiculous fish dug deeper, and debunked it:
>http://ridiculousfish.com/blog/archives/2006/05/16/36/
>
>The performance gap hinges on a single default mmap treshold parameter, which is
>set at 15Kb on MacOSX, and 128Kb on Linux, so Jasjeet's matrixes of 35Kb that are
>constantly allocated and de-allocated, were allocated from the heap with Linux,
>but had to go thru the virtual memory allocator under MacOSX.
>
>Why this parameter is set so low with MacOSX is the real question. This parameter can be changed.
>
>Argument-passing (messages) overhead was not the culprit.

I thought it would be something like that.

Many of the internal Darwin functions, those in the BSD server and libc are quite old. For years Apples developers have been focusing elsewhere. They seem to have looked at some bits of it recently though, they've improved malloc for example.

The number may be correctly tuned off-course for best performance over a large variety of loads. It's hard to tell.

Message-passing overhead is normally shown in programs that stress the OS in some way by calling it often, or making it do complex things. Not maths apps.