One of the promises of container orchestration technologies is that they enable effective utilization of hardware resources and thus reduce infrastructure costs. However, as applications scale up and down over time in a Kubernetes-managed environment, the cluster may enter a state of resource fragmentation. In such a state, the cluster's hardware cannot be used to its full potential because resources become locked by poor Pod to Node mappings. This paper shows that such problems are common as workload requirements approach a cluster's resource capacity. Additionally, we present an experimental analysis of directed Pod eviction as a technique to combat the issue of resource fragmentation. Our findings show that directed Pod eviction can reduce resource fragmentation and allow clusters to run a given workload using 16.7% fewer Nodes than is consistently possible using standard Kubernetes scheduling. These findings are unaffected by applying input shaking to the experimental analysis, strengthening confidence in the generality of these resource utilization improvements.