Application characteristics and scheduling in multi-tenant systems

In the public cloud, resources are shared between multiple tenants while cloud provider’s goal is to provide isolation, security and fairness in accounting. It has been shown in “Bolt: I Know What You Did Last Summer… In The Cloud” that shared environments observe unfairness. Unfairness can be used by either getting performance advantage over other tenants or causing wasting resources of other tenants either on purpose or unknowingly.

In this thesis the focus will be put first on detecting characteristics of workloads such as: CPU, disk, network usage patterns. The workload characteristics can be used to uniquely target attacks, identify application executed by other tenants by recognising common application patterns. However, usage patterns can be used to optimise scheduling of workload, detect potential tenants using resources unfairly. In the second part a PoC solution will be build to recognising which workloads are affected by unfair tenants and how to mitigate such impacts.

Method:

1. Create a set of testing workload that are going to be used in experiments.
2. Develop methods to generate workload signatures encapsulating resource usage patterns in environments where privileged access to some profiling tools (e.g., hardware counters) aren’t available.
3. Use statistical approach to recognise runs of workloads affected by antagonist tenants.
4. Propose solutions for the cases when unfairness is detected for example in Kubernetes cluster running in public cloud.