gc_ms2_tuning branch

[Nick Wellnhofer]

I just created a new branch gc_ms2_tuning. It contains a port of the 
dynamic GC threshold to the new GC code. The memory_used counter now 
also accounts for attribute and other fixed size storage and ignores 
constant objects. The old code could lead to a lot more memory being 
used than accounted for, so rather low thresholds could still work very 
well. To address this, the default dynamic GC threshold has been 
increased to 75%. Note that these changes have not been applied to the 
old GC. It still uses the old accounting.

It also turned out that the double linked lists in the new GC code are 
quite a hit on performance and memory consumption. I replaced them with 
a simple array as GC mark stack but afaics the double linked lists will 
also be used for Generational GC. So maybe this approach doesn't work out.

Here are some more notes on the GC thresholds to avoid confusion. There 
are two thresholds: a static and a dynamic one. The Parrot memory 
management counts the amount of memory newly allocated since the 
previous GC run. If this amount exceeds both the static and dynamic 
threshold, another GC run is triggered.

Currently, the static threshold of the old GC (ms) is hardcoded to 1MB. 
The static threshold of the new GC (ms2) is 256MB in trunk and 1MB in my 
new branch.

The dynamic threshold is a percentage of the total memory used directly 
after the last GC. (After a GC run we know exactly the minimum memory a 
process requires.) This allows us to adjust the timing of GC runs to the 
actual memory requirements of a process. It also roughly limits the 
percentage of memory wasted. Currently, the dynamic threshold of the old 
GC is 25% (although this is skewed, see my notes above). The new GC in 
trunk has no dynamic threshold, the new GC in my new branch uses 75%.

The dynamic threshold is a space-time tradeoff. Generally, lower values 
cause more GC runs, more execution time and less memory consumption. But 
on memory limited machines, a lower threshold can also improve general 
perfomance if it avoids swapping.

Nick