Inconsistency Detection Framework

Navigation: [ Introduction | System Overview | Publications | People | Progress ]



Conventionally, consistency control protocols, such as strong consistency protocols or optimistic consistency protocols that increase the availability while tolerate relaxed inconsistency among nodes, are predefined and deployed before the system starts to run an application. We refer to this scheme as inconsistency avoidance .

While inconsistency avoidance can be effective in a small-scale networked system, such as a small cluster, it has some drawbacks in an Internet-scale environment, such as Grid. In this environment, a strong consistency protocol can be very costly to maintain because of the membership maintenance and strict protocol enforcement cost.

On the other hand, optimistic consistency protocol relieves the costly maintenance and strict enforcement burden associated with strong consistency protocols; however, it also does not suit the large-scale distributed system because it is predefined. In an environment where many applications are deployed, providing a predefined consistency protocol can be either overkill when an application does not need that strong consistency, or insufficient when an application needs stronger ones.

IDF (Inconsistency Detection Framework)

Thus, we propose a framework to detect inconsistency in a timely manner when it occurs instead of avoiding it in the first place. We refer this as inconsistency detection.

IDF's primary advantages are:

System Overview

As an alternative to inconsistency avoidance, the inconsistency detection framework detects inconsistency among nodes in a timely manner. A logical diagram of this framework is shown in the figure below. In this framework, multiple applications share data and services through the support of the Internet-scale middleware and the inconsistencies among them are detected by the detector. Upon detection, the detector consults with the inconsistency level monitor (step 1 and step 2) before reaction is initiated. Based on the applications' semantics, if the inconsistency is tolerable, the detector does not react; otherwise, the detector informs the inconsistency resolution model to resolve this inconsistency (step 3).



Faculty: Hong Jiang
  Ying Lu
Student: Yijun Lu


We are currently investigating its implication to different applications.