Re: spvw_circ.d: performance

[Vladimir T. <vtz...@gm...>]

Hi,
On Sep 17, 2008, at 1:34 AM, Bruno Haible wrote:
>
> If you can reproducibly measure than WRITE-TO-STRING with :CIRCLE T  
> is faster
> with the reentrant get_circularities() than with the embedded-mark- 
> bit one,
> then please switch to this implementation if defined(VIRTUAL_MEMORY).
> (If !defined(VIRTUAL_MEMORY) the one with embedded mark bits is  
> better,
> since it does not require additional temporary memory.)
>
> I can well imagine the reason why it's faster: It writes to a more  
> concentrated
> memory area, thus causing less cache writebacks.
>

With attached patch to spvw_circ.d:

make check-tests    /* on G4 OSX */
CVS HEAD:
Real time: 440.88837 sec.
Run time: 311.81845 sec.
Space: 796342156 Bytes
GC: 1049, GC time: 84.52206 sec.

CVS HEAD + spvw_circ.d patch:
CL-USER> (time (tw2s 10000 10))
Real time: 423.2096 sec.
Run time: 298.95453 sec.
Space: 791897864 Bytes
GC: 1043, GC time: 78.07748 sec.

However I was not able to create reproducible test with WRITE-TO- 
STRING that shows any significant difference between the CVS and  
patched version.

Vladimir