On Wed, 17 Sep 2008, David Arturo Macias Corona wrote:
Hi David,
> It build hbvmmt.a without warnings, thanks
> Now Harbour is multithread in OS/2 ? :-)
Thank you very much for your tests.
It means that we have working MT in OS/2.
> b) speedtst.prg
> ST without HB_FM_STATISTICS_OFF and using hbrun speedtst.prg
> total application time: 53.30
> ST
> total application time: 46.77
> MT
> total application time: 135.88
> does not show too much difference between both ST but large with MT
It's caused by two reasons:
1. In OS2 builds we do not have any atomic inc/dec operations
so mutexes are used to protect references counters. It's hardly
visible in:
ST: c:=L_C -> 0.10
MT: c:=L_C -> 1.19
Because OS/2 uses GCC then we should be able to use the same
macros as for PTHREAD base MT Harbour builds for x86 machines.
I'll add them ASAP. Please test, you should find noticeable speed
improvement. Looks that mutexes are very expensive in OS/2.
2. The cost of TLS access. I cannot say yet how much is it yet
because it can be hidden by the mutexes used in reference counters
but this results:
ST: empty loops overhead = 0.14
MT: empty loops overhead = 0.23
suggest it's not such big. It can reduced by using stack macros
also for OS/2 MT builds. It can be quite easy done because OS2
uses a little bit specific method of accessing TLS data. I'll
make it in next step when I will know the results for eliminating
mutexes from references counters.
The memory memtest.prg mat not crash in MT mode due to different
memory allocations. It's nothing amazing and does not make the
memory manager provided by your C compiler better :-( Important
is that it can crash in some quite simple situations. I suggest
to make some test with alternative MM I'll commit ASAP.
Once again thank you for your tests.
best regards,
Przemek
_______________________________________________
Harbour mailing list
Harbour@harbour-project.org
http://lists.harbour-project.org/mailman/listinfo/harbour
