Sequentially consistent reads at wire speed

点击翻译按钮获取中文摘要

Modern data-storage systems are fundamentally inefficient. The main reason for this inefficiency is the data-consistency mechanism that ensures that read operations return the value written by the last successful write operation. Although every data item is replicated on multiple storage nodes, for any data item, only a single node is used to serve read requests, limiting the system's performance, efficiency, and scalability for supporting popular data. The goal of this project is to explore new designs for a highly efficient and performant consensus protocol that can utilize all replicas for serving read requests. The main enablers for the new protocol are recent advances in networking technology that introduced programmable switching devices. The new protocol will leverage this new network programmability to monitor every read and write request, determine when a write is completed and which subset of replicas acknowledged it, and detect node and switch failures. The new protocol will be able to not only utilize multiple replicas for serving read requests but also load balance read requests and avoid failed nodes.This innovative project will have a potentially wide influence on technology, including on many Canadian companies that build KV stores (e.g., Huawei, NetApp, Ericsson, and IBM) and the more than 150 Canadian data centers that use these systems. Finally, this program will provide an invaluable opportunity for training highly qualified personnel on cutting-edge networking and storage technology.