• KIPS Transactions on Computer and Communication Systems
• /
• v.11 no.4
• /
• pp.105-112
• /
• 2022

Recently, many studies using reinforcement learning-based autoscaling have been performed to make autoscaling policies that are adaptive to changes in the environment and meet specific purposes. However, training the reinforcement learning-based Horizontal Pod Autoscaler(HPA) policy in a real environment requires a lot of money and time. And it is not practical to retrain the reinforcement learning-based HPA policy from scratch every time in a real environment. In this paper, we implement a reinforcement learning-based HPA in Kubernetes, and propose a transfer leanring technique using a queuing model-based simulation to accelerate the training of a reinforcement learning-based HPA policy. Pre-training using simulation enabled training the policy through simulation experience without consuming time and resources in the real environment, and by using the transfer learning technique, the cost was reduced by about 42.6% compared to the case without transfer learning technique.

• KNOM Review
• /
• v.22 no.1
• /
• pp.1-10
• /
• 2019

The process to determine the required number of resources depends on the factors being considered. Autoscaling is one such mechanism that uses a wide range of factors to decide and is a critical process in NFV. As the networks are being shifted onto the cloud after the invention of SDN, we require better resource managers in the future. To solve this problem, we propose a solution that allows the VNFMs to autoscale the system resources depending on the factors such as overhead of hyperthreading, number of requests, execution-times for the virtual network functions. It is a known fact that the hyperthreaded virtual-cores are not fully capable of performing like the physical cores. Also, as there are different types of core having different frequencies so the process to calculate the number of cores needs to be measured accurately and precisely. The platform independency is achieved by proposing another solution in the form of a monitoring microservice, which communicates through APIs. Hence, by the use of our autoscaling application and a monitoring microservice, we enhance the resource provisioning process to meet the criteria of future networks.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.6
• /
• pp.1545-1559
• /
• 2023

In the cloud environment, microservices are implemented through Kubernetes, and these services can be expanded or reduced through the autoscaling function under Kubernetes, depending on the service request or resource usage. However, the increase in the number of nodes or distributed microservices in Kubernetes and the unpredictable autoscaling function make it very difficult for system administrators to conduct operations. Artificial Intelligence for IT Operations (AIOps) supports resource management for cloud services through AI and has attracted attention as a solution to these problems. For example, after the AI model learns the metric or log data collected in the microservice units, failures can be inferred by predicting the resources in future data. However, it is difficult to construct data sets for generating learning models because many microservices used for autoscaling generate different metrics or logs in the same timestamp. In this study, we propose a cloud data refining module and structure that collects metric or log data in a microservice environment implemented by Kubernetes; and arranges it into computing resources corresponding to each service so that AI models can learn and analogize service-specific failures. We obtained Kubernetes-based AIOps learning data through this module, and after learning the built dataset through the AI model, we verified the prediction result through the differences between the obtained and actual data.

• YAKHAK HOEJI
• /
• v.48 no.1
• /
• pp.60-64
• /
• 2004

Near-infrared (NIR) spectroscopy was used to determine rapidly and nondestructively the content of ambroxol in intact ambroxol tablets containing 30 mg (12.5％ m/m nominal concentration) by collecting NIR spectra in range 1100-1750 nm. The laboratory-made samples had 10.3∼15.9％ m/m nominal ambroxol concentration. The measurements were made by reflection using a fiber-optic probe and calibration was carried out by partial least square regression (PLSR) with autoscaling. Model validation was performed by randomly splitting the data set into calibration and validation data set (7 samples as a calibration data set and 5 samples as a validation data set). The developed NIR method gave results comparable to the known values of tablets in a laboratorial manufacturing Process, standard error of calibration (SEC) and standard error of prediction (SEP) being 0.49％ and 0.49％ m/m respectively. The method showed good accuracy and repeatability NIR spectroscopic determination in intact tablets allowed the potential use of real time monitoring for a running production process.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2018.10a
• /
• pp.63-66
• /
• 2018

컨테이너 기술은 운영체제 수준 가상화 기술 중 하나로 하드웨어 레벨 가상화 기술에 비해 인스턴스의 빠른 생성 및 종료시킬 수 있는 특성이 있다. 이러한 특성은 직업 부하에 따라 인스턴스의 빠른 생성 및 종료시킬 수 있는 특성이 있다. 이러한 특성은 작업 부하에 따라 인스턴스의 수량을 동적으로 조정하는 오토스케일링 상황에서 유리하게 작용할 수 있다. 본 논문에서는 다수의 노드를 기반으로 구성된 컨테이너 기반의 오토스케일링 환경과 가상머신 기반의 오토스케일링 환경을 성능 측면에서 비교하고 컨테이너 기반 환경에서 자원 할당의 변화가 성능에 주는 영향을 측정 및 분석한다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2012.11a
• /
• pp.280-281
• /
• 2012

클라우드 기술이 발달하면서 사용자 요구를 만족하면서도 비용을 절감하기 위해 VM 의 개수를 자동으로 조절해주는 오토스케일링 기술이 부각되었다. 하지만 어떤 VM 을 추가할 것인지는 NP-Hard Problem 으로 휴리스틱하게 풀 수밖에 없다. 따라서 사용자의 실시간으로 변하는 요구에 바로 대처하지 못할 수 있다. 사용자 요구에 실시간적으로 대처하기 위해서는 사용자가 보내는 요청의 패턴을 읽고, 앞으로 올 요청을 미리 아는 기술이 필요하다. 이에 본 논문에서는 리퀘스트 예측을 통한 오토스케일링을 가능케 하도록 구조를 제안하고자 한다.

• KIPS Transactions on Computer and Communication Systems
• /
• v.11 no.7
• /
• pp.205-216
• /
• 2022

One of the many tools used for distributed deep learning training is Kubeflow, which runs on Kubernetes, a container orchestration tool. TensorFlow jobs can be managed using the existing operator provided by Kubeflow. However, when considering the distributed deep learning training jobs based on the parameter server architecture, the scheduling policy used by the existing operator does not consider the task affinity of the distributed training job and does not provide the ability to dynamically allocate or release resources. This can lead to long job completion time and low resource utilization rate. Therefore, in this paper we proposes a new operator that efficiently schedules distributed deep learning training jobs to minimize the job completion time and increase resource utilization rate. We implemented the new operator by modifying the existing operator and conducted experiments to evaluate its performance. The experiment results showed that our scheduling policy improved the average job completion time reduction rate of up to 84% and average CPU utilization increase rate of up to 92%.

• Convergence Security Journal
• /
• v.23 no.5
• /
• pp.81-89
• /
• 2023

Kubernetes is a representative open-source software for container orchestration, playing a crucial role in monitoring and managing resources allocated to containers. As container environments become prevalent, security threats targeting containers continue to rise, with resource exhaustion attacks being a prominent example. These attacks involve distributing malicious crypto-mining software in containerized form to hijack computing resources, thereby affecting the operation of the host and other containers that share resources. Previous research has focused on detecting resource depletion attacks, so technology to respond when attacks occur is lacking. This paper proposes a reinforcement learning-based dynamic resource management framework for detecting and responding to resource exhaustion attacks and malicious containers running in Kubernetes environments. To achieve this, we define the environment's state, actions, and rewards from the perspective of responding to resource exhaustion attacks using reinforcement learning. It is expected that the proposed methodology will contribute to establishing a robust defense against resource exhaustion attacks in container environments