Resumen de Learning workload behaviour models from monitored time-series for resource estimation towards data center optimization
In recent years there has been an extraordinary growth of the demand of Cloud Computing resources executed in Data Centers. Modern Data Centers are complex systems that need management. As distributed computing systems grow, and workloads benefit from such computing environments, the management of such systems increases in complexity. The complexity of resource usage and power consumption on cloud-based applications makes the understanding of application behavior through expert examination difficult. The difficulty increases when applications are seen as "black boxes", where only external monitoring can be retrieved. Furthermore, given the different amount of scenarios and applications, automation is required. To deal with such complexity, Machine Learning methods become crucial to facilitate tasks that can be automatically learned from data.
Firstly, this thesis proposes an unsupervised learning technique to learn high level representations from workload traces. Such technique provides a fast methodology to characterize workloads as sequences of abstract phases. The learned phase representation is validated on a variety of datasets and used in an auto-scaling task where we show that it can be applied in a production environment, achieving better performance than other state-of-the-art techniques.
Secondly, this thesis proposes a neural architecture, based on Sequence-to-Sequence models, that provides the expected resource usage of applications sharing hardware resources. The proposed technique provides resource managers the ability to predict resource usage over time as well as the completion time of the running applications. The technique provides lower error predicting usage when compared with other popular Machine Learning methods.
Thirdly, this thesis proposes a technique for auto-tuning Big Data workloads from the available tunable parameters. The proposed technique gathers information from the logs of an application generating a feature descriptor that captures relevant information from the application to be tuned. Using this information we demonstrate that performance models can generalize up to a 34% better when compared with other state-of-the-art solutions. Moreover, the search time to find a suitable solution can be drastically reduced, with up to a 12x speedup and almost equal quality results as modern solutions.
These results prove that modern learning algorithms, with the right feature information, provide powerful techniques to manage resource allocation for applications running in cloud environments. This thesis demonstrates that learning algorithms allow relevant optimizations in Data Center environments, where applications are externally monitored and careful resource management is paramount to efficiently use computing resources. We propose to demonstrate this thesis in three areas that orbit around resource management in server environments
