I found a few interesting items in this chapter. First, the assertion (which I probably buy) that if you have a small, highly skilled team, you can forgo type safety to optimize memory footprint. Second, that deploying constructs (in this case, a device database) designed for a single machine onto a supercomputer can turn some seemingly inconsequential implementation decisions (such as device directory trees) into major bottlenecks. (20,000 devices? Yikes!)
Also, the C reference counting scheme for memory management used in this effort is very similar to the Objective-C memory management model.
Also, the C reference counting scheme for memory management used in this effort is very similar to the Objective-C memory management model.