DOCSLIB.ORG

Velos: One-Sided Paxos for RDMA Applications

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Velos: One-Sided Paxos for RDMA Applications Velos: One-sided Paxos for RDMA applications Rachid Guerraoui Antoine Murat Athanasios Xygkis EPFL EPFL EPFL Abstract These solutions primarily focus on increasing the through- put and lowering the latency of common case executions, Modern data centers are becoming increasingly equipped thus achieving consensus in the order of a few microseconds. with RDMA-capable NICs. These devices enable distributed Nevertheless, outside of their common case, these systems systems to rely on algorithms designed for shared memory. suffer from orders of magnitude higher failover times, rang- RDMA allows consensus to terminate within a few microsec- ing from 1ms [2] to tens or hundreds of ms [23, 25]. ond in failure-free scenarios, yet, RDMA-optimized algo- In this work we introduce a leader-based consensus algo- rithms still use expensive two-sided operations in case of fail- rithm based on the well-known Paxos [16] that is efficient ure. In this work, we present a new leader-based algorithm both during the common-case as well as during failover. Our for consensus based on Paxos that relies solely on one-sided algorithm provides comparative performance to the state-of- RDMA verbs. Our algorithm decides in a single one-sided . s RDMA operation in the common case, and changes leader the-art system Mu, i.e., it achieves consensus in under 1 9µ , making it only twice as slow as Mu. At the same time, our also in a single one-sided RDMA operation in case of fail- algorithm is 13 times faster than Mu in the event of a leader ure. We implement our algorithm in the form of an SMR s system named Velos, and we evaluated our system against failure and manages to failover in under 65µ . To achieve this, our algorithm relies on one-sided RDMA WRITEs as the state-of-the-art competitor Mu. Compared to Mu, our well as Compare & Swap (CAS), a capability that is present solution adds a small overhead of ≈ 0.6µs in failure-free ex- in RDMA NICs. ecutions and shines during failover periods during which it is 13 times faster in changing leader. The basic idea behind our algorithm is that the original Paxos algorithm contains RPCs that are simple enough to be replaced with CAS operations. The CAS operations are initi- 1 Introduction ated by the leader and are executedon the memoryof the par- ticipating consensus nodes (leader and followers). We first RDMA is becoming increasingly popular in data centers. describe a version of our algorithm for single-shot consensus, arXiv:2106.08676v1 [cs.DC] 16 Jun 2021 This networking technology is implemented in modern NICs where we provide proofs of correctness. Then, we continue and allows for Remote Direct Memory Access, where a by extending this version to a fully-fledged system, in which server can access the memory of another without involving our algorithm takes a form similar to multishot Paxos [17]. the CPU of the latter [24]. As a result, RDMA paves the way In the case of a stable leader, it decides in a single CAS op- for distributed algorithms in the shared-memory model that eration to a majority. In the event of a failure, the leader is are no longer confined within a single physical server. changed in a single additional CAS operation to a majority. Communication over RDMA takes two forms. One form The rest of this document is as follows: In section 2, we is one-sided communication which shares similar semantics present related work. In section 3, we introduce the consen- to local memory accesses: a server directly performs READs sus problem and present the well-known Paxos algorithm. In and WRITEs to the memory of a remote server without in- section 4, we explain how to transform the RPC algorithm volving the remote CPU. The other form is called two-sided into a CAS-based one and we prove the correctness of this communication that has similar semantics to message pass- transformation. In section 5, we discuss further practical ing: the server SENDs a message to the remote server which considerations, which are relevant in converting our single- involves its CPU to process the RECEIVEd message. shot consensus algorithm to a multi-shot one. In section 6, A significant body of work on consensus [4] over RDMA we discuss our implementation. In section 7, we evaluate has been conducted over the past decade [2, 14, 23, 25]. the performance of our solution against Mu, the most recent 1 state-of-the-art system that implements consensus. 3.2 Consensus In the consensus problem, processes propose individual val- 2 Related work ues and eventually irrevocably decide on one of them. For- mally, Consensus has the following properties: Mu Uniform agreement If processes i and j decide val and val′, Mu [2] is an RDMA-powered State Machine Replication [3] respectively, then val = val′. system that operates at the microsecond scale. Similarly to our system, it decides in a single amortized RDMA RTT. Validity If some process decides val, then val is the inputof This is achieved by relying on RDMA permissions [24]. Mu some process. ensures that at any time, at most one process can write to Integrity No process decides twice. a majority and decide, which ensures consensus’ safety. In case of leader failure, Mu requires permission changes that Termination Every correct process that proposes eventually take ≈ 250µs. Mu thus fails at guaranteeing microsecond decides. decisions in case of failure. It is well known that consensus is impossible in the asyn- chronous model [9]. To circumvent this impossibility, an ad- APUS ditional synchrony assumption has to be made. Our consen- APUS [25] is a Paxos-based SMR system. It was tailored for sus algorithm provides safety in the asynchronousmodel and RDMA. It doesn’t use expensive permission changes but re- requires partial synchrony for liveness. For pedagogical rea- lies heavily on two-sided communication schemes. While it sons and in order to facilitate understanding, we implement provides short failovers, its consensus engine involves heavy our consensus algorithm by merging together the following CPU usage at replicas and is significantly slower than Mu’s. abstractions: • Abortable Consensus [4], an abstraction weaker than Disk Paxos Consensus that is solvable in the asynchronous model, Disk Paxos [10] observes that Paxos’ acceptors can be re- • Eventually Perfect Leader Election [6], which relies on placed by moderately smart and failure-prone shared mem- the weakest failure detector required to solve Consen- ories. Their work can be done by proposers as long as they sus. are able to run atomic operations at the shared resource. This work is purely theoretical. 3.3 Abortable Consensus 3 Preliminaries Abortable consensus has the following properties: Uniform agreement If processes i and j decide val and val′, In this section, we state the process and communication respectively, then val = val′. model that we assume for the rest of this work. Then we formally introduce the problem of consensus and we present Validity If some process decides val, then val is the inputof the well-known Paxos algorithm. some process. Termination Every correct process that proposes eventually 3.1 Assumptions decides or abort. We consider the message-and-memory (M&M) model [1], Decision If a single process proposes infinitely many time, which allows processes to use both message-passing and it eventually decides. shared-memory. Communication is assumed to be lossless and provides FIFO semantics. The system has n processes Algorithm 1 solves Abortable Consensus and is based on Π = {p1,..., pn} that can attain the roles of proposer or ac- Paxos. Processes are divided into two groups: proposers or ceptor. In the system, we assume that there are p proposers acceptors. Proposers propose a value for decision and accep- and n acceptors, where 1 < p < |Π|. Processes can fail by tors accept some proposed values. Once a value has been n−1 crashing. Up to p − 1 proposers and 2 acceptors may accepted by a majority of acceptors, it is decided by its pro- fail. As long as a process is alive, its memory is remotely poser. accessible. When a process crashes, its memory also crashes. Intuitively, the algorithm is split in two phases: the Pro- In this case, subsequent memory operations do not return a pose phase and the Accept phase. During these phases, mes- response. The system is asynchronous in that it can experi- sages from the proposer are identified by a unique proposal ence arbitrary delays. number. The Prepare phase serves two purposes. First, the 2 Algorithm 1: Abortable Consensus 3.4 From Abortable Consensus to Consensus 1 Proposers execute: 2 upon hIniti: One can solve Consensus by combining Abortable Consen- 3 decided = False sus together with Eventually Perfect Leader Election (Ω). 4 proposal = id 5 proposed_value = ⊥ In Abortable Consensus a proposer is guaranteed to decide, rather than abort, if it executes unobstructed. The role of Ω 7 propose(value): 8 proposed_value = value is to ensure this condition is ensured. Informally, Ω guaran- 9 if not decided: tees that eventually all correct proposers will consider a sin- 10 if prepare(): 11 accept() gle one ofthem to be the leader. As long as a proposeris con- sidering itself as the leader it keeps on proposing their value 13 prepare(): 14 proposal = proposal + |Π| using Abortable Consensus. Eventually, Ω will mark a sin- 15 broadcast hPrepare | proposali gle correct proposer as the leader, which will try to propose 16 wait for a majority of hPrepared | ack, ap, avi 17 if any av returned, replace proposed_value with av with unobstructed and decide. The leader can then broadcast the highest ap decision to the rest of the proposers.
Load more

Recommended publications
  • 42 Paxos Made Moderately Complex
    Paxos Made Moderately Complex ROBBERT VAN RENESSE and DENIZ ALTINBUKEN, Cornell University This article explains the full reconfigurable multidecree Paxos (or multi-Paxos) protocol. Paxos is by no means a simple protocol, even though it is based on relatively simple invariants. We provide pseudocode and explain it guided by invariants. We initially avoid optimizations that complicate comprehension. Next we discuss liveness, list various optimizations that make the protocol practical, and present variants of the protocol. Categories and Subject Descriptors: C.2.4 [Computer-Communication Networks]: Distributed Syst- ems—Network operating systems; D.4.5 [Operating Systems]: Reliability—Fault-tolerance General Terms: Design, Reliability Additional Key Words and Phrases: Replicated state machines, consensus, voting ACM Reference Format: Robbert van Renesse and Deniz Altinbuken. 2015. Paxos made moderately complex. ACM Comput. Surv. 47, 3, Article 42 (February 2015), 36 pages. DOI: http://dx.doi.org/10.1145/2673577 1. INTRODUCTION Paxos [Lamport 1998] is a protocol for state machine replication in an asynchronous environment that admits crash failures. It is useful to consider the terms in this 42 sentence carefully: —A state machine consists of a collection of states, a collection of transitions between states, and a current state. A transition to a new current state happens in response to an issued operation and produces an output. Transitions from the current state to the same state are allowed and are used to model read-only operations. In a deterministic state machine, for any state and operation, the transition enabled by the operation is unique and the output is a function only of the state and the operation.
    [Show full text]
  • A General Characterization of Indulgence
    A General Characterization of Indulgence R. Guerraoui1;2 N. Lynch2 (1) School of Computer and Communication Sciences, EPFL (2) Computer Science and Arti¯cial Intelligence Laboratory, MIT Abstract. An indulgent algorithm is a distributed algorithm that, be- sides tolerating process failures, also tolerates arbitrarily long periods of instability, with an unbounded number of timing and scheduling failures. In particular, no process can take any irrevocable action based on the op- erational status, correct or failed, of other processes. This paper presents an intuitive and general characterization of indulgence. The characteri- zation can be viewed as a simple application of Murphy's law to partial runs of a distributed algorithm, in a computing model that encompasses various communication and resilience schemes. We use our characteriza- tion to establish several results about the inherent power and limitations of indulgent algorithms. 1 Introduction Indulgence The idea of indulgence is motivated by the di±culty for any process in a dis- tributed system to accurately ¯gure out, at any point of its computation, any information about which, and in what order, processes will take steps after that point. For instance, a process can usually not know if other processes have failed and stopped operating or are simply slow to signal their activity and will in- deed perform further computational steps. More generally, a process can hardly exclude any future interleaving of the processes. This uncertainty is at the heart of many impossibilities and lower bounds in distributed computing, e.g., [9], and it has been expressed in various forms and assuming speci¯c computation models,e.g., [7, 4, 19].
    [Show full text]
  • Distributed Computing Column 36 Distributed Computing: 2009 Edition
    Distributed Computing Column 36 Distributed Computing: 2009 Edition Idit Keidar Dept. of Electrical Engineering, Technion Haifa, 32000, Israel [email protected] It’s now the season for colorful reviews of 2009. While you can read elsewhere about the year in sports, top-40 pop hit charts, people of the year, and so on, I throw my share into the mix with a (biased) review of this year’s major distributed computing events. Awards First, let’s look at awards. This year we learned that two women were recognized with ACM and IEEE prestigious awards for their achievements in, (among other things), distributed computing. Barbara Liskov won the 2008 ACM Turing Award for a range of contributions to practical and theoretical foundations of programming language and system design, some of which are in distributed computing. The award citation mentions her impact on distributed programming, decentralized information flow, replicated storage, and modular upgrading of distributed systems. I include below a short reflection, by Rodrigo Rodrigues, on Barbara Liskov’s Turing Award and legacy. Looking ahead to 2010, it was recently announced that Nancy Lynch is the recipient of the prestigious 2010 IEEE Emanuel R. Piore Award1 for her contributions to foundations of distributed and concurrent computing. The award is given for outstanding contributions in the field of information processing in relation to computer science. Leslie Lamport has won the same award in 2004 for his seminal contributions to the theory and practice of concurrent programming and fault-tolerant computing. The 2009 Edsger W. Dijkstra Prize in Distributed Computing was awarded to Joseph Halpern and Yoram Moses for “Knowledge and Common Knowledge in a Distributed Environment”.
    [Show full text]
  • State Machine Replication Is More Expensive Than Consensus
    View metadata, citation and similar papers at core.ac.uk brought to you by CORE provided by Dagstuhl Research Online Publication Server State Machine Replication Is More Expensive Than Consensus Karolos Antoniadis EPFL, Lausanne, Switzerland karolos.antoniadis@epfl.ch Rachid Guerraoui EPFL, Lausanne, Switzerland rachid.guerraoui@epfl.ch Dahlia Malkhi VMware Research, Palo Alto, USA [email protected] Dragos-Adrian Seredinschi EPFL, Lausanne, Switzerland dragos-adrian.seredinschi@epfl.ch Abstract Consensus and State Machine Replication (SMR) are generally considered to be equivalent prob- lems. In certain system models, indeed, the two problems are computationally equivalent: any solution to the former problem leads to a solution to the latter, and vice versa. In this paper, we study the relation between consensus and SMR from a complexity perspect- ive. We find that, surprisingly, completing an SMR command can be more expensive than solving a consensus instance. Specifically, given a synchronous system model where every instance of consensus always terminates in constant time, completing an SMR command does not necessar- ily terminate in constant time. This result naturally extends to partially synchronous models. Besides theoretical interest, our result also corresponds to practical phenomena we identify em- pirically. We experiment with two well-known SMR implementations (Multi-Paxos and Raft) and show that, indeed, SMR is more expensive than consensus in practice. One important im- plication of our result is that – even under synchrony conditions – no SMR algorithm can ensure bounded response times. 2012 ACM Subject Classification Computing methodologies → Distributed algorithms Keywords and phrases Consensus, State machine replication, Synchronous model Digital Object Identifier 10.4230/LIPIcs.DISC.2018.7 Related Version A full version of the paper is available at [4], https://infoscience.epfl.ch/ record/256238.
    [Show full text]
  • The Impact of RDMA on Agreement
    The Impact of RDMA on Agreement Marcos K. Aguilera Naama Ben-David Rachid Guerraoui VMware CMU EPFL [email protected] [email protected] [email protected] Virendra Marathe Igor Zablotchi Oracle EPFL [email protected] [email protected] ABSTRACT provides protection mechanisms to grant and revoke access for read- Remote Direct Memory Access (RDMA) is becoming widely avail- ing and writing data. This mechanism is fine grained: an application able in data centers. This technology allows a process to directly can choose subsets of remote memory called regions to protect; it read and write the memory of a remote host, with a mechanism to can choose whether a region can be read, written, or both; and it control access permissions. In this paper, we study the fundamental can choose individual processes to be given access, where different power of these capabilities. We consider the well-known problem processes can have different accesses. Furthermore, protections are of achieving consensus despite failures, and find that RDMA can im- dynamic: they can be changed by the application over time. In this prove the inherent trade-off in distributed computing between fail- paper, we lay the groundwork for a theoretical understanding of ure resilience and performance. Specifically, we show that RDMA these RDMA capabilities, and we show that they lead to distributed allows algorithms that simultaneously achieve high resilience and algorithms that are inherently more powerful than before. high performance, while traditional algorithms had to choose one While RDMA brings additional power, it also introduces some or another. With Byzantine failures, we give an algorithm that only challenges.
    [Show full text]
  • Byzantizing Paxos by Refinement
    Byzantizing Paxos by Refinement Leslie Lamport 5 December 2010 Abstract We derive a 3f +1 process Byzantine Paxos consensus algorithm by Byzantizing a variant of the ordinary Paxos algorithm|that is, by having 2f + 1 nonfaulty processes emulate the ordinary Paxos algo- rithm despite the presence of f malicious processes. We have writ- ten a formal, machine-checked proof that the Byzantized algorithm implements the ordinary Paxos consensus algorithm under a suitable refinement mapping. Contents 1 Introduction 1 2 Consensus and Classic Paxos 2 2.1 Consensus . .2 2.2 Paxos Consensus . .2 3 Byzantizing An Algorithm 4 4 Algorithm PCon 7 5 Algorithm BPCon 8 6 Liveness and Learning About Sent Messages 10 6.1 Sending Proofs . 11 6.2 Relaying 1b Messages . 12 7 The Castro-Liskov Algorithm 13 8 The Formal Specifications and Proof 14 9 Conclusion 16 References 16 You can verb anything. Ron Ziegler (quoted by Brian Reid) 1 Introduction The Paxos algorithm [6] has become a standard tool for implementing fault- tolerant distributed systems. It uses 2f + 1 processes to tolerate the benign failure of any f of them. More recently, Castro and Liskov developed a 3f + 1 process algorithm [2] that tolerates f Byzantine (maliciously faulty) processes. Intuitively, their algorithm seems to be a Byzantine version of Paxos. Other algorithms that also seem to be Byzantine versions of Paxos have subsequently appeared [4, 11, 14]. The only previous attempt we know of to explain the relation between a Byzantine Paxos algorithm and ordinary Paxos was by Lampson [13]. He derived both from an abstract, non-distributed algorithm.
    [Show full text]
  • Table of Contents
    Table of Contents The 2010 Edsger W. Dijkstra Prize in Distributed Computing 1 Invited Lecture I: Consensus (Session la) The Power of Abstraction (Invited Lecture Abstract) 3 Barbara Liskov Fast Asynchronous Consensus with Optimal Resilience 4 Ittai Abraham, Marcos K. Aguilera, and Dahlia Malkhi Transactions (Session lb) Transactions as the Foundation of a Memory Consistency Model 20 Luke Dalessandro, Michael L. Scott, and Michael F. Spear The Cost of Privatization 35 Hagit Attiya and Eshcar Hillel A Scalable Lock-Free Universal Construction with Best Effort Transactional Hardware 50 Francois Carouge and Michael Spear Window-Based Greedy Contention Management for Transactional Memory 64 Gokarna Sharma, Brett Estrade, and Costas Busch Shared Memory Services and Concurrency (Session lc) Scalable Flat-Combining Based Synchronous Queues 79 Danny Hendler, Hai Incze, Nir Shavit, and Moran Tzafrir Fast Randomized Test-and-Set and Renaming 94 Dan Alistarh, Hagit Attiya, Seth Gilbert, Andrei Giurgiu, and Rachid Guerraoui Concurrent Computing and Shellable Complexes 109 Maurice Herlihy and Sergio Rajsbaum Bibliografische Informationen digitalisiert durch http://d-nb.info/1005454485 XII Table of Contents Brief Announcements I (Session Id) Hybrid Time-Based Transactional Memory 124 Pascal Felber, Christof Fetzer, Patrick Marlier, Martin Nowack, and Torvald Riegel Quasi-Linearizability: Relaxed Consistency for Improved Concurrency 127 Yehuda Afek, Guy Korland, and Eitan Yanovsky Fast Local-Spin Abortable Mutual Exclusion with Bounded Space 130 Hyonho
    [Show full text]
  • The Information Structure of Indulgent Consensus
    IEEE TRANSACTIONS ON COMPUTERS, VOL. 53, NO. 4, APRIL 2004 453 The Information Structure of Indulgent Consensus Rachid Guerraoui, Member, IEEE, and Michel Raynal Abstract—To solve consensus, distributed systems have to be equipped with oracles such as a failure detector, a leader capability, or a random number generator. For each oracle, various consensus algorithms have been devised. Some of these algorithms are indulgent toward their oracle in the sense that they never violate consensus safety, no matter how the underlying oracle behaves. This paper presents a simple and generic indulgent consensus algorithm that can be instantiated with any specific oracle and be as efficient as any ad hoc consensus algorithm initially devised with that oracle in mind. The key to combining genericity and efficiency is to factor out the information structure of indulgent consensus executions within a new distributed abstraction, which we call “Lambda.” Interestingly, identifying this information structure also promotes a fine-grained study of the inherent complexity of indulgent consensus. We show that instantiations of our generic algorithm with specific oracles, or combinations of them, match lower bounds on oracle-efficiency, zero-degradation, and one-step-decision. We show, however, that no leader or failure detector-based consensus algorithm can be, at the same time, zero-degrading and configuration-efficient. Moreover, we show that leader-based consensus algorithms that are oracle-efficient are inherently zero-degrading, but some failure detector-based consensus algorithms can be both oracle-efficient and configuration-efficient. These results highlight some of the fundamental trade offs underlying each oracle. Index Terms—Asynchronous distributed system, consensus, crash failure, fault tolerance, indulgent algorithm, information structure, leader oracle, modularity, random oracle, unreliable failure detector.
    [Show full text]
  • Epidemic Protocols: from Large Scale to Big Data Davide Frey
    Epidemic Protocols: From Large Scale to Big Data Davide Frey To cite this version: Davide Frey. Epidemic Protocols: From Large Scale to Big Data. Computational Engineering, Fi- nance, and Science [cs.CE]. Université De Rennes 1, 2019. tel-02375909 HAL Id: tel-02375909 https://hal.inria.fr/tel-02375909 Submitted on 22 Nov 2019 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. ANNÉE 2019 HABILITATION À DIRIGER DES RECHERCHES UNIVERSITÉ DE RENNES 1 sous le sceau de l’Université Bretagne Loire Mention : Informatique École doctorale MathStic présentée par Davide FREY préparée à l’unité de recherche Inria Rennes Institut National de Recherche en Informatique et Automatique EPI WIDE Epidemic Protocols: From Large Scale to Big Data HDR soutenue à Rennes le 11/06/2019 devant le jury composé de : SONIA BEN MOKHTAR Directeur de Recherche CNRS / Rapporteur GAEL THOMAS Professeur Telecom SudParis / Rapporteur PASCAL FELBER Professeur Université de Neuchatel / Rapporteur ANNE-MARIE KERMARREC CEO Mediego / Examinateur OLIVIER BARAIS Professeur Université de Rennes 1 / Examinateur DANNY HUGUES Professeur, KU Leuven / Examinateur Acknowledgments I would like to start by expressing my gratitude to the three reviewers, Sonia Ben Mokhtar, Pascal Felber, and Gael Thomas, for taking the time to read and and comment on this manuscript despite their busy schedules, as well as to the other members of the jury, Olivier Barais, Danny Hughes, and Anne-Marie Kermarrec.
    [Show full text]
  • Project-Team ASAP
    IN PARTNERSHIP WITH: Institut national des sciences appliquées de Rennes Université Rennes 1 Activity Report 2016 Project-Team ASAP As Scalable As Possible: Foundations of large scale dynamic distributed systems IN COLLABORATION WITH: Institut de recherche en informatique et systèmes aléatoires (IRISA) RESEARCH CENTER Rennes - Bretagne-Atlantique THEME Distributed Systems and middleware Table of contents 1. Members :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 1 2. Overall Objectives :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 2 2.1. Decentralized personalization2 2.2. Scalability: Cloud computing meets p2p2 2.3. Privacy-aware decentralized computations3 2.4. Information dissemination over social networks3 2.5. Computability and efficiency of distributed systems4 3. Research Program :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 5 3.1. Theory of distributed systems5 3.2. Peer-to-peer overlay networks5 3.3. Epidemic protocols5 3.4. Malicious process behaviors6 3.5. Online social networks and recommender systems6 4. Highlights of the Year ::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 6 5. New Software and Platforms :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 7 5.1. Asapknn (MediEgo)7 5.2. Brow2Brow 7 5.3. Dashboard 7 5.4. GossipLib 8 5.5. HEAP 8 5.6. HyRec 8 5.7. WebGC 8 5.8. WhatsUp 9 5.9. YALPS 9 5.10. P2P-AllYours 10 6. New Results ::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 10 6.1. Theory of Distributed Systems 10 6.1.1. t-Resilient Immediate Snapshot is Impossible 10 6.1.2. Two-Bit Messages are Sufficient to Implement Atomic Read/Write Registers in Crash- Prone Systems 10 6.2. Network and Graph Algorithms 11 6.2.1. Vertex Coloring with Communication and Local Memory Constraints in Synchronous Broadcast Networks 11 6.2.2. Optimal Collision/Conflict-Free Distance-2 Coloring in Wireless Synchronous Broad- cast/Receive Tree Networks 11 6.2.3.
    [Show full text]
  • State Machine Replication Is More Expensive Than Consensus
    State Machine Replication is More Expensive than Consensus Karolos Antoniadis EPFL Rachid Guerraoui EPFL Dahlia Malkhi VMware Research Group Dragos-Adrian Seredinschi EPFL Abstract Consensus and State Machine Replication (SMR) are generally considered to be equivalent problems. In certain system models, indeed, the two problems are computationally equivalent: any solution to the former problem leads to a solution to the latter, and vice versa. In this paper, we study the relation between consensus and SMR from a complexity per- spective. We find that, surprisingly, completing an SMR command can be more expensive than solving a consensus instance. Specifically, given a synchronous system model where every in- stance of consensus always terminates in constant time, completing an SMR command does not necessarily terminate in constant time. This result naturally extends to partially synchronous models. Besides theoretical interest, our result also corresponds to practical phenomena we iden- tify empirically. We experiment with two well-known SMR implementations (Multi-Paxos and Raft) and show that, indeed, SMR is more expensive than consensus in practice. One impor- tant implication of our result is that—even under synchrony conditions—no SMR algorithm can ensure bounded response times. 1 Introduction Consensus is a fundamental problem in distributed computing. In this problem, a set of distributed processes need to reach agreement on a single value [32]. Solving consensus is one step away from implementing State Machine Replication (SMR) [31, 49]. Essentially, SMR consists of replicating a sequence of commands—often known as a log—on a set of processes which replicate the same state machine. These commands represent the ordered input to the state machine.
    [Show full text]
  • Project-Team ASAP
    IN PARTNERSHIP WITH: Institut national des sciences appliquées de Rennes Université Rennes 1 Activity Report 2014 Project-Team ASAP As Scalable As Possible: foundations of large scale dynamic distributed systems IN COLLABORATION WITH: Institut de recherche en informatique et systèmes aléatoires (IRISA) RESEARCH CENTER Rennes - Bretagne-Atlantique THEME Distributed Systems and middleware Table of contents 1. Members :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 1 2. Overall Objectives :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 2 2.1. General objectives2 2.1.1. Scalability2 2.1.2. Personalization2 2.1.3. Uncertainty2 2.1.4. Malicious behaviors and privacy2 2.2. Structure of the team3 2.2.1. Objective 1: Decentralized personalization3 2.2.2. Objective 2: Personalization: Cloud computing meets p2p3 2.2.3. Objective 3: Privacy-aware decentralized computations4 2.2.4. Objective 4: Information dissemination over social networks4 2.2.5. Objective 5: Computability and efficiency of distributed systems5 3. Research Program :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 6 3.1. Distributed computing6 3.2. Theory of distributed systems6 3.3. Peer-to-peer overlay networks6 3.4. Epidemic protocols7 3.5. Malicious process behaviors7 3.6. Online social networks and recommender systems7 4. Application Domains ::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::8 5. New Software and Platforms :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 8 5.1. MediEgo: A recommendation solution for webmasters8 5.2. AllYours-P2P: A distributed news recommender (former WhatsUp)8 5.3. HyRec: A hybrid recommender system9 5.4. GossipLib: A library for gossip-based applications9 5.5. YALPS: A library for p2p applications 10 5.6. HEAP: Heterogeneity-aware gossip protocol 10 5.7. Brow2Brow: Browser-to-browser serverless toolboxes 10 5.8. WebGC: Web-based Gossip Communication 11 6. New Results ::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 11 6.1.
    [Show full text]