Improving Network Performance Through Endpoint Diagnosis

Опубликовано 28 марта 2017, 22:47

"Components of networks, and by extension the internet can fail. Users of the network (clients) do not have access to the various components of this distributed system. Therefore, it is important to provide them with tools so that they are able to handle such failures. This can be achieved in two ways. First, by helping the
client find the cause of these failures to contact the right entity and to assit in the failure resolution process. Second, by avoiding the failures and going around them. In my thesis I explore both these approaches, however, this talk will focus on the first of these two solutions. The proposed solution is based on a two step approach where the end point (without requiring ANY support from the network or the remote end hosts) identifies the entity responsible for failures, if the network is determined as the cause of the failure a second algorithm is triggered to identify the device responsible for the failure. In order to acheive this goal, we revisit the question: how much can you infer about a failure using TCP statistics collected at one of the endpoints
in a connection? Our results show that the information needed to identify the cause of performance problems in a datacenter can be gathered from end hosts effectively with low overhead while providing accuracies as high as 90%. "

See more on this video at microsoft.com/en-us/research/v...