Scheduling time-bounded dynamic software adaptation

Abstract

Component-based software increasingly needs dynamic adaptation to support applications in domains such as automotive, avionics or robotic systems. Dynamic software adaptation involves both the integration of new, previously unanticipated features and the update of existing features without requiring system downtime. Software adaptations must often be time-bounded, e.g., due to mobility constraints. Inconsistent or inaccurate behaviour may result from an adaptation that does not complete within specified time constraints. Service providers must therefore take time constraints into account when scheduling adaptation actions. This paper describes an algorithm for time-bounded scheduling of adaptation actions and demonstrates the validity of its results.

On estimating actuation delays in elastic computing systems

“…The entire dataset is available in our companion Web site. 10 We start by analyzing a Scale Out operation that deployed a new processing node to Cassandra. Figure 5 shows the evolution of the average CPU usage and CPU usage while performing disk I/O.…”

“…We can explain this behavior by recalling that Cassandra stops servicing requests when nodes are removed until it reorganizes the data. We associated the steady state after the drop to the end of the reconfiguration, 10 http://www.infosys.tuwien.ac.at/prototypes/elasticTesting/index.html therefore we considered the last change point that was detected as end of the reconfiguration. The average CPU usage had only a little increase after the reconfiguration because in the default settings Cassandra randomly distributed the load across processing nodes, and in this experiment the processing node that was removed was lightly loaded.…”

“…For example, Ramirez et al [24] propose the use of Genetic Algorithm to generate control strategies that account for actuation delay of the different control actions; Fritsch et al [10], [11] propose a time-bounded scheduling algorithm for control actions in the automotive domain; and Gambi et al [12] propose a novel technique based on model predictive control that accounts for delayed effects of scaling actions in the cloud. Consistently, all of these researchers acknowledge the importance of actuation delays, but assume that their values are known and constant.…”