RealAllocator::allocate() is inefficient because it scans each bit of the heap.

[aarontabor]

To find a free contiguous chunk of heap, the RealAllocator currently traverses the heap. This is less than optimal for several reasons:

• we know that objects have a minimum size
• we know something about alignment

By scanning the heap bit by bit, we are performing an unnecessary amount of work.

[aarontabor]

A decision has been made to use card tables to increase the efficiency of heap allocation.

[tmbasa]

I fixed this in commit e9608dc

[aarontabor]

I think this fix introduces a new error, when I run a simulation with the default collector, I'm getting seg-faults. Running the same simulation using the state of the project before this commit works as expected.

[konsna]

Please test thoroughly before updating the master branch. Work in progress can be pushed to a development branch.

[aarontabor]

It turns out that the fix e9608dc was depending on an updated version of tracefiles that have not yet been released to the entire team. When I used the previous version of the tracefiles, I was getting segmentation faults.

Since it is the previous version of tracefiles which are currently available to the team, I have reverted the changes introduced by this fix on the master branch. I have created a new branch: dev-travefileSizeInvariantFix where this fix, and other modifications required for the new version of tracefiles can be made. Once the new version of tracefiles become publicly available to the team, dev-tracefileSizeInvariantFix can be merged back into master.

[tmbasa]

The updated tracefiles are actually available on Mazder's team folder. The updated tracefiles have the number of pointers (N) adjusted to fit the object size. But these are the post-processed ones.

[tmbasa]

Modified using card table class instead of direct checking (of bitmap).

32ab5b4

[tmbasa]

Also in update: 32ab5b4,

added -pg on all subdir.mk and the makefile to allow profiling (gprof)