Whatever is locally best is not necessarily globally best. For instance, the fragment in the example might only be executed once at start up, in which case it should be ignored when considering data layout performance.
Execution profiles must be used to guide the datapath structure and RAM layout. A simple profile will give the relative execution frequency of each operator in the source code. When balancing loads a more complex profile based on the critical path through the value flow graph of the program needs to be used.
The details of data layout in memory are generally very important, especially for DRAM that is non-uniform in access time owing to its banks, rows and its cache lines when cached. Even for a single array in the user's input code, a permutation of the address lines used by the indexing subscripts may well lead to quite a big difference in performance given banked and non-uniform underlying storage. Fortunately, programmers for high-performance scientific computing (big data) are aware of this since the effects of caches and so on are just as pronounced in pure software implementations.
| 36: (C) 2012-18, DJ Greaves, University of Cambridge, Computer Laboratory. |