[omniORB] Performance

[Duncan Grisby]

On Friday 10 February, janssen.h at atlas.de wrote:

> I'm still unsatisfied with the transfer rate of omniORB. Therefore I
> have used a profiler to look behind the scenes and I find out that the
> function timedwait (in file nt.cc) costs 50% of the runtime. Another 40%
> are spent on the functions marshalChar and unmarshalChar (cdrstream.h).

Is your profiler measuring real time or processor time?  I suspect real
time, because it's normal for several threads to be blocked in
timedwait, not actually using any CPU time.

What does your IDL look like?  If marshalChar and unmarshalChar appear
in your profile, that suggests you are sending lots of individual chars,
which is particularly inefficient due to the way code set conversion has
to work.

Cheers,

Duncan.

-- 
 -- Duncan Grisby         --
  -- duncan at grisby.org     --
   -- http://www.grisby.org --