Skip to content

Latest commit

 

History

History
29 lines (19 loc) · 2.87 KB

2018.02.18-SpecPaxos.md

File metadata and controls

29 lines (19 loc) · 2.87 KB
Raw

Designing Distributed Systems Using Approximate Synchrony in Data Center Networks

Dan R. K. Ports, Jialin Li, Vincent Liu, Naveen Kr. Sharma, and Arvind Krishnamurthy

Summary

In this paper, the authors present a new network-level primitive called Mostly-Ordered Multicast (MOM) and the Speculative Paxos replication protocol, which utilize the fact that data centers provide approximate synchrony. For MOM part, the authors first explore why existing multicast solutions still have ordering violations. Then, they describe three detail design options. After that, they evaluate the ordering properties and effectiveness on the testbed and simulator. For Speculative Paxos part, they describe the model, design the protocol, and further prove the correctness of Speculative Paxos. Speculative Paxos uses a combination of techniques to eliminate packet reordering in a data center. As well, discussion and evaluation of Speculative Paxos are also presented.

Contributions

  • This paper proposes a mostly-ordered multicast (MOM) mechanism to provide best-effort ordering property for concurrent multicast operations in a data center.
  • The authors examine why existing multicast mechanisms do not provide mostly-ordered multicast property and then demonstrate how to implement MOM mechanism in the data center by presenting three design options: (1) topology-aware multicast, (2) high-priority multicast, and (3) in-network serialization.
  • The authors design and implement Speculative Paxos, which relies on MOM's ordering guarantees to efficiently sequence requests, allowing it to execute and commit client operations with the minimum possible latency and with significantly higher throughput than Paxos.
  • They evaluate the performance of Speculative Paxos and compare it to Paxos and Fast Paxos on a twelve-switch testbed. Besides, they also explore the protocol's sensitivity to network conditions by introducing MOM ordering violations.

Strengths

  • In this paper, the authors utilize the fact that data center networks are more predictable, reliable, and extensible. Thus, they co-design distributed systems and the network they use, which brings performance benefits.
  • MOM only relies on simple but effective techniques, which leverage the structured topology of a data center network and the forwarding flexibility of software-defined networking (SDN).
  • In this paper, their customized network-level multicast mechanism can achieve ordering property with greater than 99.9% probability in a data center.
  • Speculative Paxos, which is built on MOM primitive, offers 40% lower latency and 2.6x higher throughput than leader-based Paxos.

Weaknesses

  • Speculative Paxos is very sensitive to out-of-order messages, and the throughput decreases quickly with increasing number of out-of-order messages.
  • Speculative Paxos has limited application scenarios because it relies on the assumption of MOM.