CSR: Small: Scalable Transactional Replication: Theory, Protocols, and Middleware Systems

随着基于web的网络应用程序的指数级增长，它们的后端IT系统必须处理不断增长的数据量和服务请求。当应用程序工作负载生成对共享数据的并发访问时，获得高可伸缩性是一项挑战，复制共享数据以确保在出现故障时数据存活和服务可用性。解决这个问题的经典事务技术——状态机复制——是不可扩展的：规范分布式事务的提交需要解决共识，而共识的领导者是一个重要的可伸缩性瓶颈。无领导的共识协议只会在可能的情况下提供快速决策，从而徒劳无功地付出大量法定人数的代价。为了克服这些限制，该项目正在开发两种互补的技术，用于为事务系统构建可扩展的共识协议。在第一种技术中，称为凯撒方法，共识决策总是在两个通信延迟中做出，即快速决策，使用基于建议位置的方案：在一个节点上激活的事务，即事务的协调器，在事务协调器建议的位置在所有节点上执行，并在执行任何其他冲突事务之后，选择在较小的位置。要实现这一点，事务的协调器只需要知道提议的位置没有被节点的快速仲裁拒绝。然而，通过利用网络延迟和时钟漂移估计，以某种方式调整位置，使它们永远不会被拒绝。因此，为了利用快速决策，使用比解决共识所需的人数更大的快速群体的成本是通过始终以这种方式决策的能力来平摊的。在第二种技术中，称为M^2Paxos方法，事务的顺序通常仅在两次通信延迟中决定，依赖于解决共识（即大多数节点）所严格必需的经典仲裁大小。这是通过利用应用程序的数据访问局部性实现的。特别是，在低争用的情况下，M^2Paxos检查提交的事务要访问的数据，并确定负责对事务排序的节点。这允许访问相同数据的所有事务由同一节点隐式排序。该项目正在将Caesar和M^2Paxos转换为实验性的开源HyFlow事务中间件系统，这使得研究社区能够广泛采用这些技术。此外，该项目正在将这些技术转换到Red Hat/JBoss的生产事务中间件Infinispan中，从而使J2EE开发人员能够广泛采用这些技术。

With the exponentially increasing popularity of web-based networked applications, their back-end IT systems must process an ever growing volume of data and service requests. Obtaining high scalability is challenging when application workloads generate concurrent accesses on shared data that is replicated to ensure data survival and service availability in the presence of failures. The classical transactional technology for solving this problem -- State Machine Replication -- does not scale: regulating the commits of distributed transactions requires solving consensus, whose leader is a significant scalability bottleneck. Leaderless consensus protocols unfruitfully pay the cost of large quorums for providing fast decisions only whenever possible. To overcome these limitations, the project is developing two complimentary techniques for building scalable consensus protocols for transactional systems. In the first technique, called the Caesar approach, consensus decisions are always made in two communication delays, i.e., fast decisions, using a scheme based on proposed positions: a transaction activated on a node, i.e., the transaction's coordinator, is executed on all nodes at a position proposed by the transaction's coordinator, and after the execution of any other conflicting transaction that was chosen at a lesser position. To achieve that, the transaction's coordinator only needs to know that the proposed position is not rejected by a fast quorum of nodes. However, by exploiting network delays and clock drift estimates, the positions are adjusted in a way such that they are never rejected. Thus, the cost of using fast quorums larger than the ones necessary to solve consensus in order to exploit a fast decision is amortized by the ability of always deciding in that way. In the second technique, called the M^2Paxos approach, the order of transactions is generally decided in only two communication delays by relying on the classical quorum size that is strictly necessary to solve consensus, i.e., a majority of nodes. This is achieved by exploiting application's data access locality. In particular, in case of low contention, M^2Paxos inspects the data to be accessed by a submitted transaction and determines the node responsible for ordering the transaction. This allows all transactions accessing the same data to be implicitly ordered by the same node.The project is transitioning Caesar and M^2Paxos into the experimental, open-source HyFlow transactional middleware system, which enables adoption of the techniques by the research community at large. Additionally, the project is transitioning the techniques into Red Hat/JBoss's production transactional middleware, Infinispan, which enables adoption of the techniques by J2EE developers at large.