Toward Automatic Failure Diagnosis in the Cloud

点击翻译按钮获取中文摘要

As software systems have grown in size, complexity, and cost, it has become increasingly difficult to deliver bullet-proof software, resulting in many software failures in the production environment. This unfortunate fact is further exacerbated by the trend of cloud-computing - where it is even harder to make systems failure-free as they become increasingly more complex and distributed. Once a failure occurs in these production systems (i.e., users cannot receive expected service), vendors need to trouble-shoot it as quickly as possible since every minute of the downtime is costly. For example, it is estimated that Amazon will lose 2 million dollars for every hour of its downtime. Such downtime will be even more catastrophic for those mission-critical software services. For example, the 2003 northeast blackout, caused by a software bug, resulted in over 10 million people in Ontario and 45 million in U.S. out of power for at least 7 hours. Later an investigation report attributes "failure to provide effective real-time diagnostic support" as one of the main reasons behind such great damage. Finally, a prolonged trouble-shooting process will frustrate the users and significantly erode the vendor's reputation. The goal of this proposed research is to expedite the trouble-shooting of failures in cloud-based distributed systems. In particular, my research consists three progressive thrusts that together, will automate the diagnosis of such failures, and significantly reduce the downtime of the cloud-based distributed systems.