Conclusions and Future Studies
Workflow fragmentation and scheduling are significant problems in workflow management. Based on the definition by Wei Tan et al., fragmentation is a partition of workflow model. Models proposed for fragmentation have disadvantages such as increased number of generated fragments after fragmentation, which increases communication messages, time delay, and mean response time, thereby reducing throughput. Other disadvantages include static fragment generation which decreases flexibility and efficiency. Thus, the present study proposed a model for fragmentation which dynamically fragmented scientific workflows, considering runtime conditions. Moreover, it resolved the noted problems by controlling fragment generation. In WSADF framework, the number of tasks in each fragment was calculated based on the number of virtual machines. Fragments were generated during the execution, reducing communication messages among the fragments. In this study, WSADF framework was compared with the FPD algorithm in the fragmentation phase, and with FPD, CTC, Centralized, SLV, and QDA algorithms in the scheduling phase. According to the results of the experiments, response time and throughput were improved compared to the baseline studies. As the result of decreasing the number of generated fragments compared to the baseline studies, the number of communication messages among the fragments as well as delay time was reduced in this study, thereby decreasing response time and enhancing throughput. Furthermore, the results of the experiments for bandwidth usage cost and memory cost revealed the improved performance of the proposed framework compared to the baseline studies because the former controlled the number of generated fragments and selected appropriate virtual machines with less cost during runtime. Experiments were conducted in three Configuration and in both phases of fragmentor and scheduler. Results were improved compared to the baseline studies. For instance, compared to Montage workflow in Configuration-1 and the fragmentor phase, it showed 84.75% improvement in mean response time and 87.68% improvement in throughput. In Configuration2 and the fragmentor phase, it showed 84.64% improvement in mean response time and 83.46% improvement in throughput. In Configuration-1 and the scheduler phase, it demonstrated 83.94%, 69.56%, and 96.91% improvement in mean response time, throughput, and bandwidth usage cost, respectively. In Configuration-2 and the scheduler phase, it demonstrated 94.49%, 47.82%, and 96.1% improvement in mean response time, throughput, and bandwidth usage cost, respectively.
