Reviewing Cloud Monitoring: Towards Cloud Resource Profiling

Cloud data centres with providers on physical level and customers on virtual level both monitor their hard-and software infrastructure to understand load patterns and to detect malfunctions and bottlenecks. The motivation for cloud monitoring on both the virtual and physical level can be summarized to the three occasions alerting, resource allocation, and visualization. Typical cloud monitoring solutions transfer and store all metrics of all systems under observation in central stores like time series databases. Applications then query, aggregate and compute their result out of these monitoring data. In large data centres, the amount of data scales up and leads to a reasonable overhead. In addition monitoring on virtual and physical level duplicates the overhead. We present an approach for monitoring resource statistics on the physical level only, and provide resource utilisation profiles to cloud middleware and customers, instead of storing the raw time series data. The approach first revisits the necessary metrics for hardware independent resource profiles, considering overbooked physical servers as well. A profile consists of a static (e.g. CPU cores) and dynamic (e.g. changing utilisation) part, and is based on statistical computations like histograms and Markov chains.