Workflows are a set of tasks and the dependency among them, which are divided into scientific and business categories. To avoid problems of centralized execution of workflows, they are broken into segments that is known as fragmentation. To fragment the workflow, it is highly important to consider the dependency among tasks and runtime conditions. The cooperation between the scheduler and fragmentor must be such that the latter generates appropriate tasks with optimized communication cost, delay time, response time, and throughput. To this end, in the present study, a framework is proposed for scheduling and fragmentation of tasks in scientific workflows that are conducted in fragmentation and scheduling phases. In the fragmentation phase, the fragments are generated with regard to the number of virtual machines available during runtime. In the scheduling phase, the virtual machines are selected with the aim of reducing bandwidth usage. The experiments are performed with three Configurations during both phases of fragmentation and scheduling. Response time, throughput, and cost (BW and RAM) were improved compared to the baseline studies on Sipht, Inspiral, Epigenomics, Montage, and CyberShake workflows as datasets.

Introduction

Scientific workflows might be very large, carrying a vast number of tasks and calculations, manipulate a huge amount of data and they will eventually be realized as thousands of concurrent process instances. Hence, the implementation of workflows in the cloud matters [1, 2]. The centralized execution of workflows increases response time and missing the deadline of a workflows, either. As a solution, distributed workflow engines come to action; therefore, workflows are fragmented into sub-workflows (fragments) so that they can be executed using distributed resources by a cloud scheduler. Fragmentation of workflows enhances scalability and reusability, as well. Fragmentation is performed in dynamic and static modes. In the former, fragments are generated during the runtime, while in the latter, fragmentation occurs before the runtime. In static fragmentation, workflow is fragmented before the runtime and is then executed during the runtime by determining the resource. Although this method is simple, fragments are not compatible with the runtime conditions. Dynamic fragmentation decides on the generation and execution of fragments during the runtime. However, adaptable and dynamic fragmentation generates and executes the fragments based on the feedbacks from runtime environment in order to balance the scalability and efficiency of workflows. This approach is recommended when workflow engine acts as a cloud service to execute a huge number of workflow instances and/or when workflows include a large number of tasks [3]. So far, numerous studies have examined the fragmentation and scheduling of workflows. The FPD model [4] is a frequently used method in fragmentation, in which workflows are divided into single-task fragments, thereby increasing communication messages, delay time, and response time, and decreasing throughput. As a result, it is highly important to consider a method for reducing the number of fragments generated from a workflow. In the method proposed in [5], a method known as ATSDS was expressed for workflow fragmentation, which is a two-phase adaptive method for the fragmentation and scheduling of workflows during the runtime. The phases utilized in this method are fragmentation and resource allocation. Fragmentation is conducted based on the hierarchical process decentralization (HPD) algorithm that focuses on business processes and has no idea on scientific processes.