PaperlayerView exact revision on arXiv

Source-linked AI summary

Pushing BitTorrent Locality to the Limit

Stevens Le Blond, Arnaud Legout, Walid Dabbous

arXiv:0812.0581v2cs.NI

TL;DR

P2P locality may reduce inter-ISP traffic, but how far it can be pushed and its Internet-scale savings remain open questions. The paper tests high locality with large controlled BitTorrent experiments and real-torrent data. It estimates that locality would have saved 40% of inter-ISP traffic across the crawled torrents.

  • Problem

    The paper asks how far BitTorrent locality can be pushed and what inter-ISP traffic savings it can achieve at Internet scale.

    Source: §1 · p. 1

  • Method

    The authors run controlled BitTorrent experiments with up to 10 000 clients and introduce Round Robin and Partition Merging to make few inter-ISP connections feasible.

    Source: §8 · p. 13

  • Results

    40% of inter-ISP traffic—4.6 petabytes of 11.6 petabytes—would have been saved across 214 443 crawled torrents.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors report that BitTorrent performs extremely well with high locality but do not advocate using very few inter-ISP connections in real deployments.

    Source: §8 · p. 13

    The experiments do not consider a hierarchy of transit ISPs.

    Source: §3.1.2 · p. 5

Abstract

from arXiv · show

Peer-to-peer (P2P) locality has recently raised a lot of interest in the community. Indeed, whereas P2P content distribution enables financial savings for the content providers, it dramatically increases the traffic on inter-ISP links. To solve this issue, the idea to keep a fraction of the P2P traffic local to each ISP was introduced a few years ago. Since then, P2P solutions exploiting locality have been introduced. However, several fundamental issues on locality still need to be explored. In particular, how far can we push locality, and what is, at the scale of the Internet, the reduction of traffic that can be achieved with locality? In this paper, we perform extensive experiments on a controlled environment with up to 10 000 BitTorrent clients to evaluate the impact of high locality on inter-ISP links traffic and peers download completion time. We introduce two simple mechanisms that make high locality possible in challenging scenarios and we show that we save up to several orders of magnitude inter-ISP traffic compared to traditional locality without adversely impacting peers download completion time. In addition, we crawled 214 443 torrents representing 6 113 224 unique peers spread among 9 605 ASes. We show that whereas the torrents we crawled generated 11.6 petabytes of inter-ISP traffic, our locality policy implemented for all torrents would have reduced the global inter-ISP traffic by 40%.

1. INTRODUCTION

The paper asks how far BitTorrent locality can be pushed and how much inter-ISP traffic it can save at Internet scale. Large-scale experiments and a crawl of real torrents address those questions.

1. INTRODUCTION

  • Earlier locality work demonstrated benefits, but left open how far locality could be pushed without harming robustness and what savings it could achieve across the Internet.Previous estimates included indirect measurements, partial ISP measurements, and simulations; Xie et al. reported P4P savings for a single ISP.Source: §1 · pp. 1–2
  • The study uses controlled BitTorrent experiments with up to 10 000 real clients and data crawled from 214 443 torrents, 6 113 224 unique peers, and 9 605 ASes.The authors note that large-scale Internet traffic savings and their effect on customers matter to ISPs and content providers considering P2P distribution.Source: §1 · p. 2
  • High locality reduced inter-ISP traffic by up to several orders of magnitude while keeping peers’ download completion time low.The reported impact was not significantly affected by torrent size, peers per ISP, or churn; the authors also proposed strategies for challenging scenarios.Source: §1 · p. 2
  • 40% of inter-ISP traffic—4.6 petabytes of 11.6 petabytes—would have been saved across 214 443 crawled torrents spanning 9 605 ASes.The paper contrasts this Internet-scale estimate with Xie et al., whose reported reduction concerned one ISP and one large torrent.Source: §1 · p. 2

2. LOCALITY POLICY

The policy limits outgoing inter-ISP connections while adding strategies to address uneven ISP sizes and partitions. It is an evaluation implementation, not a definitive deployment proposal.

2. LOCALITY POLICY

  • The policy is a simple evaluation implementation rather than a definitive deployment solution; the authors identify two strategies to consider in implementations.Those strategies are intended to improve the efficiency and robustness of locality in challenging scenarios.Source: §2 · p. 2

2.1 Implementation of the Locality Policy

  • The tracker maps peers to regions and returns same-ISP peers, adding an outside peer only while the ISP’s outgoing-connection limit permits.The tracker approximates each ISP’s outgoing connections using peer lists it has returned, and randomizes selection among the ISP’s peers; the initial seed uses regular BitTorrent selection.Source: §2.1 · p. 2

2.2 Round Robin Strategy

  • Round Robin selects an ISP before selecting a peer, making each ISP’s selection probability independent of its size.Random selection among all outside peers instead gives larger ISPs a higher probability of receiving incoming connections; the strategies are equivalent when ISPs have equal peer counts.Source: §2.2 · p. 3

2.3 Partition Merging Strategy

  • Partition Merging addresses partitions by having a leecher request an outside peer when its neighbors have no needed pieces.The client monitors HAVE messages and uses randomized detection timers with exponential backoff; the study sets T0 and T1 to one minute. Dummy HAVE messages could hinder detection, and variants were left for future work.Source: §2.3 · p. 3

2.4 Granularity of the Notion of Locality

  • The policy can localize traffic to any network region that the tracker can identify and map peers into.The paper uses ASes rather than ISPs for its real-world scenarios.Source: §2.4 · p. 3

3. METHODOLOGY

The methodology uses large-scale experiments with a real BitTorrent client and a virtual inter-ISP topology. It evaluates traffic overhead and peer slowdown across modeled network conditions.

3.1 Experimental Setup

  • The study uses experiments rather than simulations to evaluate locality’s impact on inter-ISP traffic and BitTorrent download completion time.The authors cite simulation scale and fidelity challenges, and report behaviors visible only in real experiments with up to 10 000 peers.Source: §3.1 · pp. 3–4

3.1.1 Platform

  • Experiments run on a dedicated 178-node platform with an instrumented mainline BitTorrent 4.0.2 client; the client does not implement peer-gossip discovery such as PEX.Default clients upload at 20 kB/s to four peers and have a maximum peer set of 80; locality-aware PEX is beyond the study’s scope.Source: §3.1.1 · p. 4

3.1.2 Inter-ISPs Topology

  • The virtual topology models stub ISPs as either a full mesh of peering links or as connected to one transit ISP, with peers only in stub ISPs.ISP assignments are mapped offline, and the authors argue the experiments approximate real links except for latency; bottlenecks are also tested because they may change BitTorrent behavior. A hierarchy of transit ISPs is not modeled.Source: §3.1.2 · pp. 4–5

3.2 Evaluation Metrics

  • The study measures content replication overhead, the 95th percentile of overhead, and peer slowdown.Overhead is cross-ISP uploaded data normalized by content size; the 95th percentile is calculated from five-minute periods, and slowdown compares experimental completion time with an ideal completion time.Source: §3.2 · p. 5

4. IMPACT OF THE NUMBER OF INTER-ISP CONNECTIONS

This section tests how reducing inter-ISP connections affects traffic overhead and peer slowdown across different seed and peer-capacity scenarios. Four outgoing connections per ISP achieve the minimum overhead, while slowdown remains low except for a slow-seed case.

4.1 Experimental Parameters

  • Experiments used 1,000 peers across 10 ISPs, comparing slow-seed homogeneous, fast-seed homogeneous, and heterogeneous-capacity scenarios.The heterogeneous scenario used leechers uploading at 20, 50, or 100 kB/s and a 100 kB/s initial seed; it was included to test capacity-related effects and distinguish them from seed speed.Source: §4.1 · pp. 5–6

4.2 Impact on Overhead

  • At 4 outgoing inter-ISP connections, overhead is close to 1 versus close to 90 with BitTorrent at 3,600 connections, a reduction of up to two orders of magnitude.Four outgoing connections give the minimum attainable overhead; the reported comparison covers all three scenarios.Source: Figure 1 · p. 6, §4.2 · p. 6
  • The 95th percentile also decreases significantly with fewer outgoing connections, but is much larger in the heterogeneous scenario than in the two homogeneous scenarios.Two thirds of heterogeneous-scenario peers upload faster than the 20 kB/s peers in the homogeneous scenarios, affecting the 95th percentile used to charge transit-link traffic.Source: Figure 2 · p. 6, §4.2 · p. 6

4.3 Impact on Slowdown

  • With a fast seed, reducing connections has no significant slowdown impact; with a slow seed, four outgoing connections increase slowdown by at most 43% versus BitTorrent.The reported increase in the slow-seed scenario is attributed to poor piece diversity; the authors describe it as the worst case.Source: §4.3 · pp. 6–7, Figure 3 · p. 7

5. EVALUATION OF 4 OUTGOING INTER-ISP CONNECTIONS

With four outgoing inter-ISP connections, the policy keeps overhead near its optimum across varied torrent and network conditions, while slowdown is generally close to BitTorrent’s and rises by at most about 30% in some cases. Congested links can reverse that slowdown trade-off.

5. EVALUATION OF 4 OUTGOING INTER-ISP CONNECTIONS

  • The experiments use four outgoing connections, 100 MB content, homogeneous peer upload capacities, and a slow seed, varying one parameter at a time.The varied parameters are torrent size, peers per ISP, and inter-ISP link capacity.Source: §5 · p. 7

5.1 Impact of the Torrent Size

  • For torrents of 100 peers, overhead savings are 60%; for 10 000 peers, savings are 99.8%, while slowdown is 8% higher for 100 peers and 32% higher for 1 000 and 10 000 peers.With four outgoing connections, overhead is close to one and almost independent of torrent size.Source: §5.1 · pp. 7–8

5.2 Impact of the Number of Peers per ISP

  • With 10 000 peers, overhead remains close to 1 across peer-per-ISP settings, while slowdown is close to BitTorrent’s at 10 and 5 000 peers per ISP and around 30% higher at 100 and 1 000.The 100-peers-per-ISP case creates partitions; Partition Merging resolves the issue, and those results are used for that case.Source: §5.2 · p. 8

5.3 Impact of the Inter-ISP Link Capacity

  • With 4 outgoing inter-ISP connections, overhead remains close to 1.5 across link capacities; below 400 kB/s, BitTorrent is slower, and at 40 kB/s, this scenario is more than 200% faster.Source: §5.3 · pp. 8–9

5.4 Discussion

  • Overhead stays close to 1 in most tested scenarios and is nearly independent of torrent size, peers per ISP, and inter-ISP congestion; slowdown is close to BitTorrent’s in most cases.In some scenarios, overhead is around 30% higher than with BitTorrent.Source: §5.4 · p. 9
  • Four outgoing connections reduce overhead by up to three orders of magnitude for 10 000-peer torrents, at the cost of at most a 30% slowdown increase in these experiments.The authors characterize this slowdown increase as the price paid for the overhead reduction.Source: §5.4 · p. 9

6. REAL WORLD SCENARIOS

Real-torrent measurements and experiments show that high locality can sharply reduce inter-AS traffic while keeping download slowdown low. The Internet-scale estimate projects substantial savings across the crawled torrents, subject to the crawl and traffic-model assumptions.

6. REAL WORLD SCENARIOS

  • The authors treat ASes as the locality boundary because IP-to-AS mapping is available, while noting that ISP locality may be defined at another scale.They describe this as a coarse approximation of potential benefits.Source: §6 · p. 9

6.1 Description of the BitTorrent Crawler

  • The crawler found 6,113,224 unique peers across 214,443 torrents; its Mininova-only sample, with few Asian torrents, makes the estimated benefit a lower bound.The authors say the small fraction of fake peers is unlikely to significantly bias results.Source: §6.1 · p. 10

6.2 Impact of Locality for a Real Scenario

  • The three reference torrents span worldwide interest, local interest, and a middle-sized case, with 9,844, 4,819, and 996 peers, respectively.Experiments used torrent 1's real AS and peer distribution; the other two torrents led to the same conclusions.Source: §6.2 · p. 10, §6.2.1 · p. 10

6.2.1 Evaluation of ASes with Heterogeneous Number of Peers

  • PM+RR gives the best locality-policy slowdown, close to BitTorrent, while overhead remains up to two orders of magnitude lower.RR reduces overhead but increases slowdown for the largest ASes; PM repairs partitions, at a small overhead increase.Source: Figure 7 · p. 10, §6.2.1 · p. 11, Figure 8 · p. 11

6.2.2 Evaluation of Churn

  • With churn periods of 60 or 6,000 seconds, four outgoing inter-AS connections cut maximum overhead by one order of magnitude without worsening slowdown.The authors also report no significant change to their conclusions with a 600-second churn period, 10 to 1,000 homogeneous ASes, and other arrival patterns.Source: Figure 9 · p. 11, §6.2.2 · p. 12

6.3 Estimation of Locality Benefits at the Scale of the Internet

  • For the Internet-scale estimate, the authors assume simultaneous peer downloads and equal exchange probability, then apply a simple traffic model and reference-torrent overhead savings by AS size.They report that the model matches measured overhead for the three reference torrents; savings depend on peers per AS, not torrent size.Source: §6.3 · p. 12
  • The 100 largest torrents generate 26% of inter-AS traffic, while the 10,000 largest generate 82%; the 50 largest ASes account for 45%.ASes with the largest inter-AS traffic also receive the most significant reductions.Source: §6.3 · pp. 12–13
  • 11.6 petabytes under BitTorrent fell to 7.3 petabytes with four outgoing inter-AS connections, or 7 petabytes with PM+RR, across the crawled torrents.The PM+RR estimate is 35% above the ideal policy; the overall reduction is reported as up to 40%, or 4.6 petabytes.Source: §6.3 · pp. 12–13, Figure 11 · p. 13

7. RELATED WORK

Prior work studied locality through traces, simulations, small experiments, and a P4P field test. This paper distinguishes its evaluation by examining more parameters and real torrents at Internet scale.

7. RELATED WORK

  • Earlier studies explored locality architectures, P4P cooperation, ISP-supplied neighbor oracles, CDN-based clustering, and deterministic locality, using traces, simulations, or smaller evaluations.Reported evaluations included 45 Gnutella nodes, 120,000 Ono users, and a 700-peer simulation scenario.Source: §7 · p. 13
  • Ono reported up to 207% improvement in average peer download completion time, but no explicit inter-ISP traffic reduction.Its reported evidence included reduced IP- and AS-hop path lengths.Source: §7 · p. 13
  • The P4P field test reported around 60% inter-ISP savings for one ISP and one large torrent; this paper evaluates 214,443 torrents across 9,605 ASes and reports 40% savings using four connections.The paper describes its study as the first extensive evaluation of key parameters using large-scale experiments and real-world data.Source: §7 · p. 13

8. DISCUSSION

The paper reports that a small number of inter-ISP connections can reduce overhead while keeping slowdown low across the evaluated scenarios. It does not advocate such a small number for real deployments, and notes that realistic upload-capacity distributions remain untested.

8. DISCUSSION

  • Experiments with up to 10,000 clients and real-torrent characteristics found that a small number of inter-ISP connections reduced overhead and kept slowdown low across the evaluated parameters.The authors introduce Round Robin and Partition Merging to make small connection counts feasible for real torrents.Source: §8 · p. 13
  • The authors do not advocate such a small number of inter-ISP connections for real deployments; they aim to show that BitTorrent can still perform well with high locality.Their reported high-locality reduction was up to 40% on the crawled torrents, or 4.6 petabytes.Source: §8 · p. 13
  • Upload-capacity tests used three classes spread uniformly across peers, not a realistic capacity distribution; the authors say this is difficult to obtain at Internet scale.They suggest real deployment would better explore the impact of actual peer heterogeneity.Source: §8 · p. 13
Loading 0812.0581v2…