On Thu, 25 Sep 2008, David Arturo Macias Corona wrote:
Hi David,
> I see many surprises:
> speedtst.prg
> Harbour
> =======
> ST
> total application time: 47.13
> MT
> total application time: 51.98
>
> xHarbour
> ========
> ST
> total application time: 71.52
> MT
> total application time: 119.68
Your results are more similar to the ones I have in my Linux box
and only Harbour MT overhead is bigger.
I took current Harbour and xHarbour code and unmodified speedtst.prg
from Harbour SVN and I have:
Harbour
ST: total application time: 25.10
MT: total application time: 27.99
xHarbour:
ST: total application time: 38.23
MT: total application time: 51.95
Harbour MT overhead to ST mode is: 11.51%
xHarbour MT overhead to ST mode is: 35.89%
xHarbour ST mode is 52.31% slower then Harbour
xHarbour MT mode is 86.60% slower then Harbour
In your results:
Harbour MT overhead to ST mode is: 10.27%
xHarbour MT overhead to ST mode is: 67.34%
xHarbour ST mode is 51.75% slower then Harbour
xHarbour MT mode is 130.26% slower then Harbour
As I said the only one noticeable difference is bigger MT overhead
in xHarbour for OS2 then in xHarbour for Linux. The rest is in practice
proportionally the same.
The overhead ~10-15 percent seems to be minimal cost we have to pay for
MT mode in Harbour and cannot be significantly reduced.
I also made test for cost of MT overhead in Linux when native compiler
TLS support is disabled in Harbour.
MT: total application time: 37.58
The overhead is huge: 49.72% and this is more close to the results
in Windows which Viktor has.
I made BCC build in my linux box and the results are:
ST: 34.32
MT: 42.61
MT ooverhead is 24.16%
Now we have HB_STACK_TLS_PRELOAD macro which allows to use
Mindaugas inlined assembler code for TLS access:
ST: 34.21
MT: 41.60
It reduces the overhead to 21.60%. Rather small. Any how I'll
add it to SVN code so users can make test in real windows environment.
If they do not show noticeable improvement then I'll remove it from
SVN.
Finally I tested cross build version of MinGW-3.4.5 and I'm very
surprised with the results:
ST: 22.35
MT: 31.69
The ST code is faster then native linux code. I noticed that programs
executed by WINE runs fater on real multi CPU machine but I haven't
expected anything like that.
The MT overhead is 41.79%.
This MinGW version does not have native TLS support so I create asm
inline function based on the one created for BCC by Mindaugas and it
reduced the MT speedtst time to 28.88 sec. (overhead 29.22%).
I'll commit both inline asm functions to SF repository in a while
and I would like to ask BCC and MinGW users to make speed tests for
MT mode for Harbour compiled with and without -DHB_NO_TLS macro.
It will be the answer if it's worth to keep them or not.
best regards,
Przemek
_______________________________________________
Harbour mailing list
Harbour@harbour-project.org
http://lists.harbour-project.org/mailman/listinfo/harbour
