No abstract

Abstract-This paper proposes Clock-RSM, a new state machine replication protocol that uses loosely synchronized physical clocks to totally order commands for geo-replicated services. Clock-RSM assumes realistic non-uniform latencies among replicas located at different data centers. It provides low-latency linearizable replication by overlapping 1) logging a command at a majority of replicas, 2) determining the stable order of the command from the farthest replica, and 3) notifying the commit of the command to all replicas. We evaluate Clock-RSM analytically and derive the expected command replication latency. We also evaluate the protocol experimentally using a geo-replicated key-value store deployed across multiple Amazon EC2 data centers.

Many current online services are deployed over geographically distributed sites (i.e., datacenters). Such distributed services call for geo-replicated storage, that is, storage distributed and replicated among many sites. Geographical distribution and replication can improve locality and availability of a service. Locality is achieved by moving data closer to the users. High availability is attained by replicating data in multiple servers and sites. This paper considers a class of scalable replicated storage systems based on deferred update replication with transactional properties. The paper discusses different ways to deploy scalable deferred update replication in geographically distributed systems, considers the implications of these deployments on user-perceived latency, and proposes solutions. Our results are substantiated by a series of microbenchmarks and a social network application.

Many database replication protocols are based on the deferred update replication technique. In deferred update replication, transactions are executed by a single server, and certified and possibly committed by every server. Thus, servers must store a full copy of the database. This assumption is detrimental to performance since servers may not be able to cache the entire database in main memory. This paper introduces RAM-DUR, a variation of deferred update replication whereby transaction execution is in-memory only. RAM-DUR's key insight is a sophisticated distributed cache mechanism that provides high performance and strong consistency without the limitations of existing solutions (e.g., no single server must have enough memory to cache the entire database). In addition to presenting RAM-DUR, we detail its implementation, and provide an extensive analysis of its performance.

This paper explores the possibility of implementing the widely deployed Paxos consensus protocol in network devices. We present two different approaches: (i) a detailed design description for implementing the full Paxos logic in SDN switches, which identifies a sufficient set of required OpenFlow extensions; and (ii) an alternative, optimistic protocol which can be implemented without changes to the OpenFlow API, but relies on assumptions about how the network orders messages. Although neither of these protocols can be fully implemented without changes to the underlying switch firmware, we argue that such changes are feasible in existing hardware. Moreover, we present an evaluation that suggests that moving Paxos logic into the network would yield significant performance benefits for distributed applications.