Quantification of Sequential Consistency in Actor-like Systems: An Exploratory Study

For sequentially-trained programmers, sequential consistency, i.e. program operations run in the same order as they appear in code, is the most intuitive consistency model to understand their pro-grams. Recently variations of the actor model have been added to programming languages and libraries as a concurrency mechanism. Actor models, in general, do not guarantee sequential consistency. A surprising observation, studied in detail here using a large (130 KLOC) set of benchmarks, is that: the variation of the actor model supported by a language or library causes sequential inconsisten-cies exhibited by programs to vary greatly. Knowing the impact of these variations on sequential inconsistencies is important for fo-cusing testing and verification efforts: for instance, if the variation supports in-order messaging then programs have 53 % less sequen-tial inconsistencies; or support for data isolation allows triggering of 75 % of sequential inconsistencies by only controlling interleav-ings of 2 actors and 2 messages. 1