Re: Tclkit vs Tclkit Lite VFS performance

On 13 Apr 2007, at 10:10, Pat Thoyts wrote:

[...]
> tclkit stat : 151.223
>   lite stat : 3809.82
> tclkit mtime : 153.387
>   lite mtime : 3792.815
> tclkit glob : 1739.74
>   lite glob : 36003.36
> 154.94 microseconds per iteration
> called Tcl_Close on channel with refCount > 0
[...]
> There is a catastophic slowdown
> in the 8.5 case and I will guess from  previous experience that this
> is to do with the new arbitrary length integer code. MD5 had trouble
> with this at one point although I don't know if these tests run
> something like that.

Hm, good to know.  Can't think of a place where MP arith code might  
kick in.  Or could it be TclVFS, with its 64-bit file sizes and seek  
offsets?

> I don't know at this time what is causing the stacktrace for 8.5.
> Possibly some rechan based issue.

With 8.5, and the new chan command, it might be time to jettison r(e) 
chan.  I don't know how much work it'd be to implement a memchan  
equivalent in pure Tcl, but it sure would be nice to get rid of the  
rechan hack.  The less special code tclkit has, the better - and  
apart from zlib and a small amount of startup code in kitInit.c,  
rechan is really the last thing that remains.

I don't see this extreme slowdown with 8.5 on Mac OS X, btw.  Looks  
like the above performance problem can only be debugged on Windows.

-jcw