Efficient gossip and robust distributed computation

Chryssis Georgiou, Dariusz R. Kowalski, Alex A. Shvartsman

Research output: Contribution to journalArticle

Abstract

This paper presents an efficient deterministic gossip algorithm for p synchronous, crash-prone, message-passing processors. The algorithm has time complexity T = O(log2 p) and message complexity M = O(p1+ε), for any ε > 0. This substantially improves the message complexity of the previous best algorithm that has M = O(p1.77), while maintaining the same time complexity. The strength of the new algorithm is demonstrated by constructing a deterministic algorithm for performing n tasks in this distributed setting. Previous solutions used coordinator or check-pointing approaches, immediately incurring a work penalty Ω(n + f·p) for f crashes, or relied on strong communication primitives, such as reliable broadcast, or had work too close to the trivial Θ(p·n) bound of oblivious algorithms. The new algorithm uses p crash-prone processors to perform n similar and idempotent tasks so long as one processor remains active. The work of the algorithm is W = O(n + p·min{f + 1, log3 p}) and its message complexity is M = O(fpε + pmin{f + 1, log p}), for any ε > 0. This substantially improves the work complexity of previous solutions using simple point-to-point messaging, while "meeting or beating" the corresponding message complexity bounds. The new algorithms use communication graphs and permutations with certain combinatorial properties that are shown to exist. The algorithms are correct for any permutations, and in particular, the same expected bounds can be achieved using random permutations.

Original languageEnglish (US)
Pages (from-to)224-238
Number of pages15
JournalLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume2848
StatePublished - Dec 1 2003
Externally publishedYes

Fingerprint

Gossip
Distributed Computation
Message Complexity
Crash
Time Complexity
Permutation
Checkpointing
Random Permutation
Algorithm Complexity
Deterministic Algorithm
Message Passing
Idempotent
Broadcast
Communication
Immediately
Message passing
Penalty
Trivial
Graph in graph theory

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Cite this

Efficient gossip and robust distributed computation. / Georgiou, Chryssis; Kowalski, Dariusz R.; Shvartsman, Alex A.

In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 2848, 01.12.2003, p. 224-238.

Research output: Contribution to journalArticle

@article{8b8a7c5d331e4af6ac3eaf8f67daaff1,
title = "Efficient gossip and robust distributed computation",
abstract = "This paper presents an efficient deterministic gossip algorithm for p synchronous, crash-prone, message-passing processors. The algorithm has time complexity T = O(log2 p) and message complexity M = O(p1+ε), for any ε > 0. This substantially improves the message complexity of the previous best algorithm that has M = O(p1.77), while maintaining the same time complexity. The strength of the new algorithm is demonstrated by constructing a deterministic algorithm for performing n tasks in this distributed setting. Previous solutions used coordinator or check-pointing approaches, immediately incurring a work penalty Ω(n + f·p) for f crashes, or relied on strong communication primitives, such as reliable broadcast, or had work too close to the trivial Θ(p·n) bound of oblivious algorithms. The new algorithm uses p crash-prone processors to perform n similar and idempotent tasks so long as one processor remains active. The work of the algorithm is W = O(n + p·min{f + 1, log3 p}) and its message complexity is M = O(fpε + pmin{f + 1, log p}), for any ε > 0. This substantially improves the work complexity of previous solutions using simple point-to-point messaging, while {"}meeting or beating{"} the corresponding message complexity bounds. The new algorithms use communication graphs and permutations with certain combinatorial properties that are shown to exist. The algorithms are correct for any permutations, and in particular, the same expected bounds can be achieved using random permutations.",
author = "Chryssis Georgiou and Kowalski, {Dariusz R.} and Shvartsman, {Alex A.}",
year = "2003",
month = "12",
day = "1",
language = "English (US)",
volume = "2848",
pages = "224--238",
journal = "Lecture Notes in Computer Science",
issn = "0302-9743",
publisher = "Springer Verlag",

}

TY - JOUR

T1 - Efficient gossip and robust distributed computation

AU - Georgiou, Chryssis

AU - Kowalski, Dariusz R.

AU - Shvartsman, Alex A.

PY - 2003/12/1

Y1 - 2003/12/1

N2 - This paper presents an efficient deterministic gossip algorithm for p synchronous, crash-prone, message-passing processors. The algorithm has time complexity T = O(log2 p) and message complexity M = O(p1+ε), for any ε > 0. This substantially improves the message complexity of the previous best algorithm that has M = O(p1.77), while maintaining the same time complexity. The strength of the new algorithm is demonstrated by constructing a deterministic algorithm for performing n tasks in this distributed setting. Previous solutions used coordinator or check-pointing approaches, immediately incurring a work penalty Ω(n + f·p) for f crashes, or relied on strong communication primitives, such as reliable broadcast, or had work too close to the trivial Θ(p·n) bound of oblivious algorithms. The new algorithm uses p crash-prone processors to perform n similar and idempotent tasks so long as one processor remains active. The work of the algorithm is W = O(n + p·min{f + 1, log3 p}) and its message complexity is M = O(fpε + pmin{f + 1, log p}), for any ε > 0. This substantially improves the work complexity of previous solutions using simple point-to-point messaging, while "meeting or beating" the corresponding message complexity bounds. The new algorithms use communication graphs and permutations with certain combinatorial properties that are shown to exist. The algorithms are correct for any permutations, and in particular, the same expected bounds can be achieved using random permutations.

AB - This paper presents an efficient deterministic gossip algorithm for p synchronous, crash-prone, message-passing processors. The algorithm has time complexity T = O(log2 p) and message complexity M = O(p1+ε), for any ε > 0. This substantially improves the message complexity of the previous best algorithm that has M = O(p1.77), while maintaining the same time complexity. The strength of the new algorithm is demonstrated by constructing a deterministic algorithm for performing n tasks in this distributed setting. Previous solutions used coordinator or check-pointing approaches, immediately incurring a work penalty Ω(n + f·p) for f crashes, or relied on strong communication primitives, such as reliable broadcast, or had work too close to the trivial Θ(p·n) bound of oblivious algorithms. The new algorithm uses p crash-prone processors to perform n similar and idempotent tasks so long as one processor remains active. The work of the algorithm is W = O(n + p·min{f + 1, log3 p}) and its message complexity is M = O(fpε + pmin{f + 1, log p}), for any ε > 0. This substantially improves the work complexity of previous solutions using simple point-to-point messaging, while "meeting or beating" the corresponding message complexity bounds. The new algorithms use communication graphs and permutations with certain combinatorial properties that are shown to exist. The algorithms are correct for any permutations, and in particular, the same expected bounds can be achieved using random permutations.

UR - http://www.scopus.com/inward/record.url?scp=35248829130&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=35248829130&partnerID=8YFLogxK

M3 - Article

AN - SCOPUS:35248829130

VL - 2848

SP - 224

EP - 238

JO - Lecture Notes in Computer Science

JF - Lecture Notes in Computer Science

SN - 0302-9743

ER -