PhD Proposal: Responsive Consistency in User-Centric Dynamic Network Environments

Talk
Benjamin Bengfort
Time: 
12.06.2016 15:30 to 17:00
Location: 

AVW 3450

Data replication has largely become the central technique in creating durable, highly available, multi-user systems and as a result the fault-tolerant coordination of replicas for correctness and consistency has become an increasingly important research area. As these systems have grown in scale and been deployed in geographically distributed clouds, the research focus has shifted away from networked file systems toward data-centric, cloud-managed database systems. One motivation for this shift has been the need to support an increasing number of mobile devices and another to provide scalability in collaborative workloads. However, with the advent of the Internet of Things and the increased use of multiple devices and local collaboration, we expect that local, user-centric clouds of heterogenous devices and mobile, dynamic networks will become increasingly important.

We contend that current consistency models are highly dependent on the network environment and that existing protocols can fail to achieve progress or provide sufficient consistency for user applications in many environments outside of the data center. Weak or eventual consistency models rely on the fast propagation of updates and therefore consistency is dependent entirely on message latency. Strong, sequential consistency provided by consensus protocols require multiple messages passed through a leader and though it can be shown that such protocols are correct, leadership introduces a bottleneck and partitions can make progress impossible. In user-centric dynamic clouds, responsive consistency models must be developed to take advantage of the changing network environment and respond to user requirements.

We hypothesize that we can provide responsive consistency in heterogenous, mobile topologies through two mechanisms: the federation of multiple consistency models depending on the environment and the load balancing of leadership costs in a quorum through hierarchical consensus. Federated consistency allows heterogenous devices to select a variety of consistency polices that are integrated through a core consensus group. By allowing flexibility locally, devices can make progress through consistency models that take advantage of the current network environment, but still have global guarantees. Hierarchical consensus improves the core consensus group, making it highly available and scalable by balancing leadership to specific decision spaces based on access patterns. We propose that federated consistency and hierarchical consensus together will produce a responsive consistency model that provides strong enough correctness guarantees and high enough availability to implement an effective multi-user distributed file system.

Examining Committee:

Chair: Dr. Peter Keleher

Dept rep: Dr. James Reggia

Member: Dr. Amol Deshpande