Integrated System Models for Reliable Petascale Storage Systems
The big challenges facing petascale storage systems are not actually those of performance or traditional fault tolerance schemes, but instead the challenges stem from the inherent complexity of large distributed systems. The best way to address these problems is to build a system model into the system itself, and have the system continuously monitor its current state as compared with the desired state according to the model. Change is applied to the model first, and then the system strives to achieve the new model. The hard part is defining a good model, building good support for introspection, and exploiting these features in the system design.