Strategies employed for LHC software performance studies

The objective of this work is to collect and assess the software performance related strategies employed by the major players in the LHC software arena: the four main experiments (ALICE, ATLAS, CMS and LHCb) and the two main software frameworks (Geant4 and ROOT). As the software used differs between the parties, so do the directions and methods in optimization, and their intensity. The common feeling shared by nearly all interviewed parties is that performance is not one of their top priorities and that maintaining it at a constant level is a satisfactory solution, given the resources at hand. In principle, despite some organized efforts, a less structured approach seems to be the dominant one, and opportunistic optimization prevails. Four out of six surveyed groups are investigating memory management related effects, deemed to be the primary cause of their performance issues. The most commonly used tools include Valgrind and homegrown software. All questioned groups expressed the desire for advanced tools, suitable for use by individual non-expert users, thus indirectly indicating limited will to spawn concentrated activities by experts. This paper outlines several recommendations, which (if implemented) might allow optimization efforts to be more effective.