Risk-Based Dynamic Allocation of Computing Resources

摘要

We consider the dynamic allocation of computing resources to multiple classes of workload streams with uncertainty through tradeoffs among associated risks and returns over a given time horizon. Instances of this general resource allocation problem arise naturally in many different contexts, such as those faced by cloud computing service providers. The risks and returns associated with each workload differ from one class to the next as a per-class function of the resources allocated to the class, where the service provider receives revenues for serving the class and incurs costs both for the resources that it allocates and for any service-level agreement violations due to insufficient resources. These risks and returns also evolve over time according the dynamics and uncertainty of each workload. Meanwhile, the service provider seeks to optimize its overall return-risk objectives within the context of various restrictions on resource allocations. A common objective is to maximize the expected returns over a given horizon, while maintaining the risk exposure of its resource allocation decisions, usually represented by the variability of returns, within acceptable levels. As workload dynamics and uncertainty are revealed, the decisions on resource allocations to each workload class are periodically reviewed and adjusted over time.