Rethinking State-Machine Replication for Parallelism

Abstract

Abstract-State-machine replication, a fundamental approach to designing fault-tolerant services, requires commands to be executed in the same order by all replicas. Moreover, command execution must be deterministic: each replica must produce the same output upon executing the same sequence of commands. These requirements usually result in single-threaded replicas, which hinders service performance. This paper introduces Parallel State-Machine Replication (P-SMR), a new approach to parallelism in state-machine replication. P-SMR scales better than previous proposals since no component plays a centralizing role in the execution of independent commands-those that can be executed concurrently, as defined by the service. The paper introduces P-SMR, describes a "commodified architecture" to implement it, and compares its performance to other proposals using a key-value store and a networked file system.

Archie

“…Recently, a redesign of the state-machine approach, called P-SMR, has been proposed in [16]. The motivation of both Archie and P-SMR is similar: both want to increase the parallelism in state-machine replication.…”

“…Archie only assumes the existence of transactional read and write operations to instrument. As all DER-based systems, snapshot-deterministic [18,16] transactions are assumed. This excludes any forms of non-determinism.…”

“…For the sake of generality and following the trend of [13,10,16], we adopt the programming model of software transactional memory (STM) [24] and its natural extension to distributed systems (i.e., DTM). DTM allows the programmer to simply mark a set of operations with transactional requirements as an "atomic block".…”

“…One approach that overcomes this limitation consists of relaxing the total order while providing a set of partial orders where only conflicting transactions are commonly serialized [16]. This allows the parallelization of the serial phase and also of the ordering process but requires additional information from the application to classify submitted transactions.…”

“…This way, they do not process transactions but simply broadcast transaction requests to all nodes and wait until the fastest replica replies. This method is known as Deferred Execution Replication (DER) [16,10].…”