Refining Microservices Placement Employing Workload Profiling Over Multiple Kubernetes Clusters

Abstract

As cloud-native computing is becoming the de-facto paradigm in the cloud field, Microservices Architecture has attracted attention from industries and researchers for agility and efficiency. Moreover, with the popularity of the IoT in the context of edge computing, cloud-native applications that utilize geographically-distributed multiple resources are emerging. In line with this trend, there is an increasing demand for microservices placement that selectively use optimal resources. However, optimal microservices placement is a significant challenge because microservices are dynamic and complex, depending on diversified workloads. Besides, generalizing workloads' characteristics consisting of complex microservices is realistically challenging. Thus, microservices deployment with mathematically structured algorithms based on simulation is less practical. As an alternative, a microservices placement framework is required that can reflect the characteristics of workloads derived from empirical profiling. Therefore, in this research work, we propose a refinement framework for profiling-based microservices placement to identify and respond to workload characteristics in a practical way. To achieve this goal, we perform profiling experiments with selected workloads to derive delicate resource requirements. Then, we perform microservices placement with a greedy-based heuristic algorithm that considers application performance by using resource requirements derived from the profiled results. Finally, we verify the proposed concept by comparing the experimental results that use our work and those that don't.

Service Placement for Latency Reduction in the Fog Using Application Profiles

“…From the selection of works highlighted in Table 1, some approaches are based on service requirements, so the service can be deployed in nodes with specific hardware characteristics [7], [8], [31]. Other proposals use profiling to determine when to scale services, after identifying workloads and saturation points [9], [32]. Empirical profiling for service placement has also been explored, although limited to placement of Virtual Machines into physical ones [10].…”

“…Likewise, there are proposals of orchestration architectures [7] that consider scheduling/planning modules based on application profiling, but limited to the study of Network Funtions Virtualization (NFV) [6]. Some works present the empirical evaluation of their proposal using testbeds including technologies such as Kubernetes [9], [31], [32]. Thus, to the best of our knowledge, there are no proposals for using application profiles in the Fog based on application features (e.g., behavior) beyond their resource requirements.…”

“…As open issues regarding the service placement of applications composed of microservices, one possibility is studying VOLUME 4, 2016 the use of hybrid metrics, combining profiles of applications according to their behavior, and network-related metrics that can guide the placement process to an optimal solution. Both metrics can change during time, adjusting themselves to reflect the current conditions of the network and the demands of users, unlike other metrics used in related works, such as hop count [28] or service resource requirement [7]- [9], [31].…”

“…Although the use of profiles has been explored in the context of the Cloud, the adoption of the microservices architecture brings new challenges, given the fine granularity of the components that can even be shared between different applications. To the best of our knowledge, the use of behavior profiles in Cloud to IoT scenarios has not been thoroughly exploited so far from the academic point of view, because of the complexity of the landscape and the lack of datasets available for study to apply data analytics techniques, with some studies using application requirement profiles [7]- [9], [31]. Nevertheless, there have been some recent efforts to gather real-time data in this kind of environment that are still in an early stage [6].…”

“…Applications can be profiled according to different factors, such as their resource requirements [7], [8] and their workload [9], [10]. Another possibility is using their popularity, measured by their request count, in order to prioritize them during their placement.…”