Research On Performance Monitoring And Adaptive Pre-replication Chkckpoint Of Docker's Fault Tolerance

As a new-type virtualized technology,Docker Container has drawn significant attention recently.It successfully improves the safety and failure recovery of Docker by isolating operating system kernel entities so as to achieve security isolation.Docker has gradually become an essential element of PaaS;however,the reliability of Docker container has developed into one of hot issues today.Moreover,container checkpoint is designed to automatically stock operation status of Docker at specific times in storage medium(aka,Docker container checkpoint operation).To be specific,it is a technology of backward recovery of fault tolerance,aiming at realizing container's high availability.Under circumstances of container cloud services,there is obvious dynamic variability in users' requirements.A high load business access can lead to malfunction of the container,thus,the container's availability at high load condition has to be well-maintained.And then container stop providing services.To achieve this goal,based on the load monitoring of the container cluster,failure recovery technology of container checkpoint will be used.Accordingly,the container's checkpoint will be able to recover when a failure occurs,and it ensures the normal operation of the container services.This study,bottomed on Docker Swarm container cluster,has firstly come up with a container performance monitoring plan at the basis of Zabbix to realize the load monitoring of the container cluster.This plan that monitors load performance of container cluster achieves to send a warning notice to users when the container is over loaded.A probabilistic prediction algorithm inspired by Markov Chain is applied on account of the fact that future loads of the container constantly change.Experiments have proved that this algorithm is capable of precisely forecasting the container's future load condition according to current load value of the container.As for time overhead of container checkpoints,a strategy has been designed that automatically adapts to prereplication container checkpoints;consequently,the downtime of container checkpoints is reduced.This strategy mainly includes two parts:prereplication and stop-copy.After several rounds of prereplication,memory pages that do not often get revised will be replicated first.Dirty pages generated during last round of prereplication will go through refrigerated containers and get conserved in every round of prereplication.Meanwhile,the number of rounds of prereplication is adaptively determined by work volume,and then the downtime of checkpoints gets lessened.The experiment results show that compared with existing methods,SPC-CKPOT succeeds in realizing shorter checkpoint downtime when memory allocation and workload continue to increase in the container.Finally,a prototype system is designed and established,improving the reliability of performance monitoring.Docker container performance data is collected and decoded.The accuracy of the probability prediction algorithm for the future load state of the container based on the Markov Chain is verified by experiments,and the effectiveness of the adaptive pre-replication container checkpoint strategy in reducing the time cost of the container checkpoint is verified.