Measuring, understanding, troubleshooting and optimizing various aspects of a Cloud Service hosted in remote data centers is a vital, but non-trivial task. Carefully arranged and analyzed periodic measurements of Cloud-Service latency can provide strong insights into the service performance. A Cloud Service may exhibit latency and jitter which may be a compound result of various components of the remote computation and intermediate communication. We present methods for automated detection and interpretation of suspicious events within the multi-dimensional latency time series obtained by CLAudit, the previously presented planetary-scale Cloud-Service evaluation tool. We validate these methods of unsupervised learning and analyze the most frequent Cloud-Service performance degradations.
展开▼