record deep dive

[Kostis Sagonas]

Ulf Wiger wrote:
 >
 > I also tried a ring benchmark with increasing payload,
 > using the hybrid heap, and it performed excellently,
 > with message passing cost of abot 1 us regardless of 
 > message size.

Yes.  The hybrid heap architecture has a lot of potential,
and we have spent some effort in making it more robust than
in previous versions, but let me stress that the hybrid heap
is still experimental and requires more tuning.  I strongly
recommend playing with it, but it is still not mature to be
used in production code.

 > Apparently, the compiler is capable of
 > realizing that the output of lists:duplicate/2 should
 > be put on the shared heap, since it is later passed 
 > as an argument in spawn_link()? Impressive.

Thanks for the "Impressive", but unfortunately this is not true.
We still have not hooked the changes to the compiler for the
hybrid heap.  This ability to recognize shared data is currently
only part of the hybrid runtime system.  As such it does indeed
help but it has a runtime cost.  Wait till you see the changes
in the compiler being hooked in... and then you'll really be
impressed !

Cheers,
Kostis.