It seems like the number of physical compute cores counted by Parallel is less than other tools (e.g., lscpu) by a factor of the number of sockets, but I haven't tested it on many architectures. In order to fix this, I think the socket/core/thread counting scheme implemented by Parallel should be changed. Specifically, the "core id" associated with a particular "physical id" should be counted, not just the unique number of "core id" values. Alternatively one could just multiply the number of cores by the number of sockets and be done, but that doesn't seem as reliable. What do you think?
-- Thanks, Joe Sapp
