Steering of sequential jobs with a distributed shared memory based model for online steering

Abstract

Online steering systems allow one to retrieve application data like intermediate results for visualization, and to modify parameters during the run-time of the application. While most steering systems use a client/server paradigm, online steering can favorably be modeled as a distributed shared memory with concurrent access by the application and the online steerer. In this paper, this idea is formalized, focusing on the exploration of the consistency models and protocols for the distributed shared memory. The behavior of the steering system is described by consistency models, which also guarantee the data integrity of the application, both within a single process and between multiple application processes. Depending on the integrity requirements, applications can choose the proper model and protocol. The performance of our protocols is evaluated with a synthetic workload, which shows that the newly developed delayed weak consistency is faster than the special weak consistency. Furthermore, the results prove that the invalidate protocols of both consistency models are able to adapt themselves to the workload.