PDF Publication Title:
Text from PDF Page: 061
CHAPTER 3. KNOWN REVISIONS 61 Once phase one is completed, we will refer to this proposer as the Leader8. The leader is the distinguished proposer and thus is responsible for reaching decisions. If another proposer suspects that the leader has failed, the proposer can take over as leader by executing phase one, we refer to this process as leader election9 and thus become the next distinguished proposer. The key advantages of Multi-Paxos is that during steady state, each decision is reached in one round trip to the majority of acceptors and one synchronous write to persistent storage. The system is in steady state when exactly one of the proposers (the leader) is in the replication phase and a majority of acceptors are up and responsive. A system should be operating in this state most of the time. Multi-Paxos places substantial load on the leader. In the steady state, this single proposer is responsible for receiving candidate values, assigning values to indexes, proposing values to acceptors, collecting accept messages, learning decided values and notifying participants of decisions. For this reason, the leader is often the bottleneck in Multi-Paxos systems. This unbalanced approach leads to high pressure on the leader and its network links, whilst leaving the other participants and other areas of the network under-utilised. Furthermore, whilst the system is now able to achieve consensus in one round trip instead of two, there is also only one proposer who can achieve this. Candidate values must therefore be forwarded to the leader (or clients redirected) which can add an additional round trip. These reasons are the motivation for algorithms such as Mencius [MJM08, §3]. 3.7 Roles So far in this thesis, we have divided the responsibilities in Classic Paxos into two distinct roles: acceptors and proposer10. This approach was adopted as it is widely used in the academic literature, however this distinction is also quite arbitrary. We could for example have only one role called replica which co-locates proposer and acceptor into one participant. The replica would benefit from a reduction in the number of acceptors it needs to communicate with by one and the proposer could use the acceptors last promised epoch when generating the next epoch. This approach is widely discussed in the academic literature and adopted in practice, exam- ples include Simple Paxos [Lam01a, §3], Chubby [CGR07], Mencius [MJM08], VRR [LC12], Raft [OO14] and Moderately Complex Paxos [VRA15, §4.4]. 8The leader is also known in the literature as master, primary [LC12] or coordinator [MPSP10]. Non- leader proposers are also known as backups [LC12] and followers [OO14, §5.1]. The term leader here should not be confused with leaders, which is sometimes used as an alternative term for proposers, for example by Renesse and Altinbuken [VRA15]. 9This is referred to view change in Viewstamped Replication [OL88, LC12] 10Proposer can be subdivided into distinguished and non-distinguished or leaders and non-leadersPDF Image | Distributed consensus
PDF Search Title:
Distributed consensusOriginal File Name Searched:
UCAM-CL-TR-935.pdfDIY PDF Search: Google It | Yahoo | Bing
Cruise Ship Reviews | Luxury Resort | Jet | Yacht | and Travel Tech More Info
Cruising Review Topics and Articles More Info
Software based on Filemaker for the travel industry More Info
The Burgenstock Resort: Reviews on CruisingReview website... More Info
Resort Reviews: World Class resorts... More Info
The Riffelalp Resort: Reviews on CruisingReview website... More Info
| CONTACT TEL: 608-238-6001 Email: greg@cruisingreview.com | RSS | AMP |