According to Christoph Badura:
> 
> I think that is the wrong approach.  The system shouldn't overcommit VM
> except when the application *explicitly* requests it.
> 
That would be a neat solution, but I'm pretty certain it will break a 
lot of applications.  Pity though, because a clearly differentiated 
malloc()/brk() would enable all the statistic gathering that is so 
sorely missed in this thread.

I guess a compromise is called for; perhaps a kernel or even an 
application-global option that indicates that overcommitment is not 
required and then special memory allocation functions to override this 
behaviour.  Where the option is not set, the alternative functions are 
equivalent.  If the option is set at kernel level, statistics can be 
gathered about overcommitment and applications terminated or notified 
(same thing, really, just a signal) accordingly.  In the other instance, 
no such facility is available and current behaviour can be expected.

++L