Automating inventory composition management for bulk purchasing cloud brokerage strategy

Abstract

Cloud providers offer end-users various pricing schemes to allow them to tailor VMs to their needs, e.g., a pay-as-you-go billing scheme, called on-demand, and a discounted contract scheme, called reserved instances. This work presents a cloud broker that offers users both the flexibility of on-demand instances and some discounts found in reserved instances. The broker employs a buy-low-and-sell-high strategy that places user requests into a resource pool of pre-purchased discounted cloud resources.

A key challenge to buy-in-bulk-sell-individually cloud broker business models is to estimate user requests accurately and then optimise the stock level accordingly. Given the complexity and variety of the cloud computing market space, the number of the regression model and inherently optimisation search space can be intricate.

In this thesis, we propose two solutions to the problem. The first solution is a risk-based decision model. The broker takes a risk-oriented approach to dynamically adjust the resource pool by analysing user request time series data. This approach does not require a training process which is useful at processing the large data stream. The broker is evaluated with high-frequency real cloud datasets from Alibaba. The results show that the overall profit of the broker is closely related to the optimal case. Additionally, the risk factors work as intended. The system hires more reserved instances when it can afford while leaning to the on-demand otherwise. We can also conclude that there is a correlation between the risk factors and the profit. On the other hand, the risk factor possesses some limitations, i.e. manual risk configuration, limited broker setting.

Secondly, we propose a broker system that utilises the concept of causal discovery. From the risk-based solution, we can see that if there are parameters correlated with the profit, then by adjusting those parameters, we can manipulate the profit. We infer a function mapping from the extracted key entities of broker data to an objective of a broker, e.g. profit. The technique is similar to the additive noise model, causal discovery method. These functions are assumed to describe an actual underlying behaviour of the profit with respect to the parameters. Similar to the risk-based, we use the Alibaba trace data to simulate long term user requests. Our results show that the system can infer the underlying interaction model between variables unlock the profit model behaviour of the broker system.