Abstract:
The key role in achieving high reliability, security, fault tolerance, and low latency of query service in distributed systems (including cloud computing) is played by the consolidation of data processing and storage resources in clusters, the efficiency of which increases with the use of virtual machine technologies and container virtualization. The complexity of building queuing models for container virtualization systems is caused by the fact that the intensity of query execution in each container is associated with the dynamic division of shared resources between active (performing functional tasks) containers and the costs of supporting all containers deployed in the VM, including inactive containers waiting for service requests to be sent to them. The reduction in service intensity in each container due to shared resource allocation depends on many factors that are difficult to investigate. For clusters with container virtualization, this article provides an approximate boundary estimate of the average request waiting time and the probability of timely service. When building an analytical model, each container is represented as a separate single-channel queuing system with an infinite queue and the simplest input stream. The key feature of the proposed virtual cluster model is the estimation of upper, lower, and average bounds for the potential service intensity reduction in containers, resulting from the allocation of a node's limited computing resources among them. This depends on the number of deployed containers and the dynamically varying count of active containers, which is influenced by the input stream intensity. The study demonstrates the existence of an optimal number of containers per node, minimizing the average request processing time or maximizing the probability of timely request execution. The proposed models can be applied to the structural and parametric optimization of clusters with pipelined virtualization, including in the case of scaling and reconfiguration adaptive to traffic changes by disconnecting or connecting some of the deployed containers depending on changes in the load in the system.
Fulltext:
PDF file (2965 kB)