Abstract
Docker introduces swarm mode to provide native support for managing a cluster of Docker engines as a swarm. But unfortunately, the default scheduler in Docker swarm does not perform well enough when services in Docker swarm try to cope with load changes. We find that many images share some common Blobs(Binary Large Objects) from same base images after surveying some most popular application images from Docker Hub. Based on the survey and the understanding that the number of replicas for a service is not greater than the scale of its instances, we develop a new scheduler, called the Availability-assured Buffered-layer Prioritized scheduler(the ABP scheduler), to quicken the creation and spread of services in Docker swarm. The ABP scheduler leverages local buffered layers on nodes in Docker swarm to reduce network traffics and accelerate startup of correlative tasks for services. Our experiments show that the ABP scheduler significantly shortens the latency of scaling out services from more than 10 seconds to less than 2 seconds and hastens service creation in Docker swarm under some circumstances.