Improve core assignment with multiple NUMA regions per socket about benchexec HOT 3 CLOSED

This is also a problem if several runs should be executed with few cores but much memory, such that for example the cores for all runs are taken from the first NUMA region, but not enough memory is available in that region.

from benchexec.

To handle multiple nodes per socket we can alter the cores_of_package to cores_of_region in resources.py.

memory_regions = []
for core in allCpus:
    coreDir = '/sys/devices/system/cpu/cpu{0}/'.format(core)
    memory_regions.append(_get_memory_banks_listed_in_dir(coreDir))

And correspondingly we build a cores_of_region mapping instead of cores_of_package

for core, region in zip(allCpus, memory_regions):
    cores_of_region[region].append(core)

Now using cores_of_region shall not affect the distribution for systems with single node per socket(or package) and for systems with multiple NUMA regions the algorithm shall split the socket into different groups each with their own different NUMA region.
This shall improve core assignment as runs shall be split accross memory regions instead of packages, thus ensuring that all the memory regions are used equally.

from benchexec.

Sounds good. Please make sure that the code would still work for systems that have multiple packages per NUMA node (we do not know such systems, but they could exist). In this case, actual packages should be used, of course (we will always use whatever is more fine granular: NUMA nodes or packages).

from benchexec.