P2P网络中用户参与动态：模型分析与应用多样性

需积分: 1 133 浏览量更新于2024-08-02 收藏 270KB PDF 举报

在《Characterizing Churn in Peer-to-Peer Networks》的技术报告中，作者们探讨了在设计和评估 Peer-to-Peer (P2P) 应用时至关重要的一个概念：用户驱动的节点参与度或"churn"。P2P系统的本质特性在于其节点的动态参与，这种动态性对系统的稳定性和效率有着显著影响。以往的研究已经指出，P2P系统中的节点参与行为非常活跃且变化无常，但缺乏详尽的模型用于模拟和深入分析。报告的核心挑战在于，虽然之前的研究揭示了节点参与的动态特性，但并未提供一个统一的模型来量化和预测不同类型的P2P应用中的这种行为。为了填补这一空白，研究人员进行了深入的研究，重点关注了三种不同类别的广泛部署P2P系统：无结构化的文件共享系统（如Gnutella）、内容分发系统（如BitTorrent）以及分布式哈希表系统（如Kad）。他们试图解答的问题是：这些系统的节点参与动态行为是否表现出一致性，或者它们之间是否存在显著差异。研究发现了一些关于P2P系统中churn的重要特征： 1. **相似性：**令人惊讶的是，尽管这些系统的具体工作原理和功能各不相同，但总体上的节点参与动态模式在很大程度上是相似的。这意味着，尽管应用类型有别，但核心的节点加入和离开行为可能存在通用规律。 2. **行为模式：**通过详细的分析，报告揭示了节点参与度随时间的变化模式，包括高峰期、低谷期以及可能的季节性波动。理解这些模式有助于设计者更好地规划资源分配和负载均衡策略。 3. **影响因素：**研究还探索了可能影响节点参与度的因素，如网络拓扑、内容吸引力、用户行为习惯以及系统更新等，这些都是决定churn的关键因素。 4. **设计启示：**了解这些系统间的共性和差异对于设计更加健壮和适应性强的P2P应用至关重要。设计师需要考虑到churn带来的挑战，如数据冗余、服务可用性和系统稳定性，并据此优化设计决策。 5. **评估框架：**该研究也为评估和比较不同P2P应用的性能提供了一个标准化的框架，使得跨应用的比较和优化成为可能。这篇论文为理解和管理P2P网络中的节点流动性提供了关键洞察，为P2P系统的持续改进和发展奠定了基础。通过深入分析和模型化，研究人员希望能够引导未来的P2P技术朝着更高效、稳定的方向发展。

. Alternatively, a peer q that was part of snapshot

(i − 1) but was not present in snapshot i must have left

during the interval 2∆ from the start of one crawl and

the end of the next. Therefore, we can measure with

a granularity of 2∆ the departure and arrival times of

every peer. We note that as the number of active peers

grows, the duration of each crawl (∆) increases, and thus

the granularity of our measurements becomes coarser,

i.e., there is a tradeoff between the size of a snapshot

and the accuracy of the measured arrival and departure

times. Therefore, it is essential to minimize the duration

of crawls (∆). Once we determine the departure and ar-

rival times of a peer p within a sequence of back-to-back

snapshots, we can easily determine the duration of one

appearance which is called its session time as follows:

SessionT ime = DepartureT ime − ArrivalT ime.

We also use the term uptime for active peer p to denote

the duration of time since its arrival.

To ensure that measured session times are not biased,

we use the “create-based method” employed by Saroiu

et al. [15]: Given a sequence of back-to-back snapshots

during a window of τ minutes, we split the measurement

window into two halves. Then, we only keep the ses-

sion time for those peers that (i) arrive during the ﬁrst

half, (ii) leave during either the ﬁrst or second half of

the measurement window, and (iii) their session time is

not longer than

. This guarantees unbiased results for

sessions shorter than

, but tells us nothing about the dis-

tribution of longer sessions. To avoid time-of-day bias in

our results we chose τ = 2 days. Our initial measure-

ments, as well as previous studies [3], show ﬂuctuations

in network size correlated with the time of day.

In the following subsections, we present a brief

overview of our candidate applications, and discuss

application-speciﬁc issues in capturing accurate and rep-

resentative snapshots.

2.1 BitTorrent

BitTorrent is a popular P2P application that is often used

for the distribution of very large ﬁles from a source to

a large group of users (called a swarm). Peers form an

overlayand exchange different blocks of the content until

each peer has the entire ﬁle. Each swarm is coordinated

by a rendezvous point, called a tracker, whose address is

provided out of band. Each new peer contacts the tracker

to join the swarm, periodically sends an update of its

progress, and informs the tracker when it departs. Note

that each peer may receive the entire ﬁle across multi-

ple sessions, i.e., it may obtain only a subset of blocks

in one session and resume the download later. Since the

tracker logs all its interactions with group members, the

The interval is 2∆ rather than ∆ because there is a possibility

the peer arrives during crawl i − 1 after the crawler has passed its

neighborhood.

log provides detailed information about the arrival and

departure times of each peer.

We have obtained tracker logs from two long Bit-

Torrent swarms: distributions of Debian and Red Hat

Close examination of these tracker logs reveals that

roughly 50% of participating peers contact the tracker

within every 5 minutes, and 99% of them contact the

tracker within every31 minutes. However, peers may de-

part in an ungraceful fashion and abruptly stop contact-

ing the tracker. To identify these peers, we conservatively

assume any peer that has not contacted the tracker within

35 minutes has ungracefully departed. These make up

around one third of all sessions in our dataset and were

eliminated since we can not measure their session time.

We note that the session time for a BitTorrent client is

a combination of time spent downloading the ﬁle (the

download time) and additional time that the user leaves

the client running after the download is complete (the

lingering time). While the download time might be in-

ﬂuenced by the size of the ﬁle or the number of other

peers, the lingering time is directly determined by user

behavior. Furthermore, the user can directly control the

duration of each session by stopping the application dur-

ing the download and returning at a later time to com-

plete the ﬁle download. Since the tracker log presents

the evolution of delivered content to each peer, it allows

us to separate download time from lingering time in our

analysis and examine them separately.

2.2 Gnutella

Gnutella is a popular P2P ﬁle-sharing applications with

more than 1.3 million concurrent peers [19]. Each peer

joins the network by connecting to a random group of

participating peers. Since Gnutella is not run as a dae-

mon, the arrival and departure times of each peer are trig-

gered by user behavior, i.e., session times are driven by

when the user opens and closes the application. There is

no central node in the Gnutella network that keeps track

of all participating peers, therefore the only way to dis-

cover all peers is to crawl the overlay. Given a few par-

ticipating peers in the session, a crawler progressively

contacts peers to learn about their neighbors, until it dis-

covers all the peers. The large size of the Gnutella net-

work makes it a challenge to capture a complete crawl

quickly. To address this, previous studies have selected a

random subset of peers discoveredby a partial crawl, and

periodically probe those peers to measure their session

time (e.g., [15, 3]). The key question is “Does the ses-

sion times of such a subset of peers represent the entire

population of sessions in the Gnutella network?”. With

a heavy-tailed distribution of session time [16, 5], peers

We would like to thank Ernst Biersack from the Institut Eurecom

who has kindly shared their Red Had tracker logs with us [7]. We

obtained the Debian tracker logs directly from the Debian organization.

剩余14页未读，继续阅读

linxiaoqin3555

粉丝: 1
资源: 13

P2P网络中用户参与动态：模型分析与应用多样性

Characterizing the Torque Lookup Table of an IPM Machine for Automotive

Pitfalls and Tradeoffs in Simultaneous, On-Chip FPGA Delay Measurement

the code of using the second moment beam width to calculate the beam radius

characterizing and avoiding negative transfer

characterizing differential amplifiers for communications circuits

给出一个关于牙齿分割的最优化实际问题，用数学语言描述并且给出python代码解决

一款纯VF控制的变频器方案方案说明:可做0.2KW7.5KW 220V，0.2KW75KW 380V，富士通MB90F462A

基于Java语言实现的软件工程Lab1-2021111888设计源码

基于Python的入门级人脸、视频、文字检测与识别项目设计源码

SANJIAO_JICHU gerber.zip

最新资源