[ntar-workers] Some perf test results on NTAR.

Guy Harris guy at alum.mit.edu
Tue Jul 12 19:26:25 GMT 2005


On Jul 12, 2005, at 12:20 PM, Gianluca Varenni wrote:

> Probably. Fortunately enough, MS ships the source of the CRT with  
> VSNET2003 (at least the sources for the static CRT), so I'll have a  
> look at how they implement that stuff.
> Another idea I have is to actually disable the FILE buffering using  
> setbuf(), and see what happens. At least on Windows, I expect an  
> improvement, because in any case the OS has its own global file  
> caching mechanism.

The same is true on UN*X systems, but it's probably still more  
expensive to copy data from a kernel-mode buffer with a read()/ 
ReadFile() than to copy it from a user-mode buffer, as you don't have  
the system call overhead or the overhead for finding various OS  
kernel data structures for the file and the page in question.  That's  
one reason why the FILE * routines exist - to avoid making a read()  
call on every read.



More information about the ntar-workers mailing list