Biblio

Cloud services have made large contributions to the agile developments and rapid revisions of various applications. However, the performance of these applications is still one of the largest concerns for developers. Although it has created many performance analysis frameworks, most of them have not been efficient for the rapid application revisions because they have required performance models, which may have had to be remodeled whenever application revisions occurred. We propose an analysis framework for diagnosis of application performance anomalies. We designed our framework so that it did not require any performance models to be efficient in rapid application revisions. That investigates the Pearson correlation and association rules between system metrics and application performance. The association rules are widely used in data-mining areas to find relations between variables in databases. We demonstrated through an experiment and testing on a real data set that our framework could select causal metrics even when the metrics were temporally correlated, which reduced the false negatives obtained from cause diagnosis. We evaluated our framework from the perspective of the expected remaining diagnostic costs of framework users. The results indicated that it was expected to reduce the diagnostic costs by 84.8\textbackslash% at most, compared with a method that only used the Pearson correlation.

Traditional Anti-virus technology is primarily based on static analysis and dynamic monitoring. However, both technologies are heavily depended on application files, which increase the risk of being attacked, wasting of time and network bandwidth. In this study, we propose a new graph-based method, through which we can preliminary detect malicious URL without application file. First, the relationship between URLs can be found through the relationship between people and URLs. Then the association rules can be mined with confidence of each frequent URLs. Secondly, the networks of URLs was built through the association rules. When the networks of URLs were finished, we clustered the date with modularity to detect communities and every community represents different types of URLs. We suppose that a URL has association with one community, then the URL is malicious probably. In our experiments, we successfully captured 82 % of malicious samples, getting a higher capture than using traditional methods.