Orchestrating Bag-of-Tasks Applications with Dynamically Spawned Tasks in a Distributed Environment

As the distributed computing paradigm continues to evolve due to advances such as cloud and fog computing, the need for effective workload allocation on distributed resources has become more imperative than ever. However, workload orchestration in distributed environments poses many challenges, especially in the case where the workload comprises jobs consisting of parallel component tasks. A particularly challenging scenario in the context of such bag-of-tasks applications, is the case where a component task produces an output that does not meet a predefined Quality of Service (QoS) threshold. In such a case, a new task is dynamically spawned at runtime, in order to carry out the additional computational work required for producing an acceptable result. In order to leverage data locality, the dynamically spawned task must be assigned to the same resource as the previous task that failed to meet the QoS requirements. Consequently, it is important to examine how a such dynamically changing workload affects the performance and load balancing of the distributed resources. To this end, in this paper we investigate the orchestration of bag-of-tasks applications with dynamically spawned tasks in a distributed environment. Three task routing techniques are studied and compared using simulation, under various load cases.

I. Introduction

Recent technological advances, such as cloud and fog computing, have contributed to the evolution and widespread adoption of the distributed computing paradigm. More and more applications are being migrated to such environments, in an attempt to harness the computational power of the distributed resources. This trend, along with the explosive growth of big data analytics facilitated by machine learning and other artificial intelligence (AI) techniques, resulted in applications that feature a high degree of parallelism, operating on large and diverse data sets. Applications of this kind typically consist of component tasks that are independent and able to be processed in any order. Such workloads are commonly referred to as bag-of-tasks (BoT) jobs [1], [2]. The execution of a BoT job is considered finished only when the processing of all of its component tasks has been completed. Examples of BoT jobs include data mining, chromosome mapping, drug discovery and image processing [3]–[5].