[gclist] Program behaviour models.

[Hans Boehm]

On Jan 15,  8:54am, Charles Fiterman wrote [after slight editing]:
> Subject: [gclist] Program behaviour models.
> Instead of having a one size fits all model for
> program behavior perhaps programs could inform
> collectors about which model is most valuable.
>
> 0. Program dominated by execution time.
> 1. Program dominated by page swap time.
> 2. Program expands to a size and keeps it.
> 3. Program expands at peak period then contracts.
>

Some of this information, particularly 1, often isn't known statically.
It depends on memory size, input size, etc.  It can also often be determined
dynamically.  The (old) Xerox PCR collector, for example, did switch marking
algorithms if it detected that the process was paging a lot.  This saved
on the order of a factor of 2 (more if you do some other things wrong)
for programs that were paging heavily.  Unfortunately, it increased the
complexity of the collector and introduced some problems as processes
approached
the swap space limit.

Hans


-- 
Hans-Juergen Boehm
boehm@mti.sgi.com