• No results found

Taming the Torrent: A Practical Approach to Reducing Cross ISP Traffic in Peer to Peer Systems David R. Choffnes and Fabian E.

N/A
N/A
Protected

Academic year: 2021

Share "Taming the Torrent: A Practical Approach to Reducing Cross ISP Traffic in Peer to Peer Systems David R. Choffnes and Fabian E."

Copied!
14
0
0

Loading.... (view fulltext now)

Full text

(1)

Taming the Torrent:

A Practical Approach to Reducing Cross‐ISP Traffic in  Peer‐to‐Peer Systems David R. Choffnes and Fabian E. Bustamante Presented by: Alan Dunn Topics in Network Protocol Design April 9, 2010

Outline

• Motivation

• Background – BitTorrent and CDNs • Approach – How to use CDNs

(2)

Motivation

• Peer‐to‐Peer (P2P) traffic is a significant  portion of all Internet traffic (70%? Cited study omits North America, Asia…) • Largest P2P network: BitTorrent • BitTorrent traffic effect on ISPs: –+$: Users acquire broadband to use BitTorrent –‐$: BitTorrent traffic is potentially costly

Motivation (cont’d)

• ISPs want way of managing cost of BitTorrent but keeping income • Throttle traffic? –Difficult to maintain, anger users –Questionable legality • Cache content? –Extra equipment cost –Also questionable legality…

(3)

Motivation (cont’d)

• Lessen traffic impact? –In particular cross‐ISP traffic – links to other ISPs  can be bottleneck, may cost • How to decide which peerings lessen cross ISP  traffic? –AS number? Some ISPs have many… –ISP‐based oracle • ISP service ranks peers by its preference • But ISPs have to run these, users have to trust them

CDN Redirection as Oracle

• Content Distribution Networks (CDNs) redirect  traffic based on connection quality • Advantages: – ISPs don’t have to commit extra resources (though get  less direct control) – Users don’t have to trust ISPs • Paper’s goal: – Show how to use CDN redirection as an oracle – Empirically evaluate that this reduces cross‐ISP traffic

(4)

Background: BitTorrent

Main architecture pieces: • Torrent – File metadata • Trackers – Distribute peer sets for files • Peers – Hold file data, transfer to each other

Background: BitTorrent (cont’d)

Tracker (1.2.3.4) music.torrent: Tracker: 1.2.3.4 Piece length: 2^18 Piece hashes: ff0123…, 789ebc…, … 1. New peer downloads torrent Peer New Peer 2 3 4 2. Peer contacts tracker 3. Tracker assigns peer set “music.torrent please” I am peer 4 music.torrent: Leechers: 2,3 Seeds: 1 Contact 1,2 4. Connect to peers “Hello” 5. (Optional) Bitmap of pieces “I have pieces  1, 2, 5, 7” 6. Requests, pieces, interest, choking music.torrent: Leechers: 2,3, 4 Seeds: 1

(5)

BitTorrent: Protocol Lessons

• Tracker keeps little information about peer  status, but controls peer knowledge of other  peers • Peers have complete control over where they  fetch pieces when they know peers exist • How to “prefer” some peers: –Ensure you are connected to them always –“Tit‐for‐tat” allocation: Ensure they have more  bandwidth available from you

Background: CDNs

• CDNs contain copies of content dispersed at  different network locations • Request for content leads to redirection to  selected CDN servers –Eg: via URL rewriting, DNS redirection

(6)

How a CDN redirection works:

Source: Drafting Behind Akamai, by Su et. al. (includes current  paper’s authors)

Background: CDNs (cont’d)

• Prior work: CDN servers (Akamai specifically)  returned track live network conditions  (minimize RTT) • PlanetLab nodes used to repeatedly query  domains served by Akamai (every 20 s) • Retrieved servers were pinged (every 5 s) • List of top ten (by lowest RTT, 9 = best, 0 =  worst) tracked over time

(7)

Background: CDNs (cont’d)

• Calculated rank of Akamai’s picks in the RTT lineup • For some nodes, Akamai picked particularly well (here  14 or more 90% of the time) • Low‐level Akamai DNS reporting to Brazilian PlanetLab node was known to have bad redirection times

How to use CDNs

• DNS query for CDN served sites • Keep track of result of DNS queries for a  specific name • Measure similarity between query results for  different BitTorrent peers

(8)

How to use CDNs (cont’d)

• Ratios per BitTorrent peer per site form a map • CDN servers cluster by /24 prefixes • Example (here using raw numbers not ratios): Peer 1 resolves a20.g.akamai.net: Results: 192.124.226.40, 192.124.226.27, 192.124.250.5, 192.124.250.1, 192.124.226.40 ‐> <192.124.226.0/24: 3, 192.124.250.0/24: 2>

How to use CDNs (cont’d)

• Similarity between peers calculated via “cosine  similarity” a.k.a. “dot product of unit vectors” Peer 2 resolves a20.g.akamai.net: ‐> <192.124.226.0/24: 4, 192.124.250.0/24: 2, 192.124.91.0/24: 1> Peer 1 resolves a20.g.akamai.net (from before): ‐> <192.124.226.0/24: 3, 192.124.250.0/24: 2> Similarity:      = (3*4 + 2*2)/(sqrt(21)*sqrt(13)) ≈ 0.968 | || |a b b a

(9)

How to use CDNs – Extra Details

• Threshold for selection: 0.15 – For reference, in 2D: angle < 81.4⁰ • Map contents account for age: – 1) Exponentially decay all entries – 2) Increase query result to where ratio sum is 1 • Adjusting query interval – Start with bootstrap phase: Every 30 s for 2 min – Increase by 1 minute each time no change – Decrease by half each time change occurs (minimum:  30 s)

How to use CDNs – Efficiency

• OK to use CDNs this way? (Not their intended  purpose…) –Low bandwidth (18 KB up, 36 KB down per client  per day) • Though low bandwidth does not necessarily imply low  load on CDNs (eg. CPU load)… –CDN selection strategies may change

(10)

Evaluation – What to measure

• How “cross‐ISP” is traffic? – Number of IP hops, AS hops – Sufficient? (Authors themselves comment on # AS  hops != # ISP changes, but then measure this  anyway…) • Route quality – Important for adoption by users – RTT, packet loss rate, throughput • Differences between different CDN sites (“quality  of different oracles”)

Evaluation

• Plugin “Ono” for BitTorrent client Azerus (now  Vuze) • Performs CDN DNS lookups –Even on behalf of non‐Ono peers to make up for  number of Ono users (adds measurement error?) • Measures throughput, latency, routes (via  traceroute) for each peer connection

(11)

Degree of Cross ISP Traffic

• Each point is avg. hop number for Ono‐ recommended or total peers over six‐hours • Single CDN site resolution used  (LeMonde.com)

Route Quality – RTT and Loss

• Seems to confirm that CDN locating correlates  with latency

(12)

• Connections in data have reached 4 KB/s at some point • Ono: “Significantly” higher peak rates (not shown) • Average dl: 31% higher, Median actually ~ 2 KB/s less  than standard • Problem is BitTorrent uses many low‐bandwidth  connections instead of few high‐bandwidth ones

Route Quality – Throughput (cont’d)

• Ono can take advantage of extra network  bandwidth when present • Loses in equal allocation case due to lower  RDSNet = Higher bandwidth to  in‐network connections EasyNet = Same bandwidth to  in‐network connections

(13)

Route Quality – Throughput (cont’d)

• Confirmed transfer rates were due to network  policies • Suggests RDSNet‐like policy for ISPs + CDN  redirection would be a winning combination for  both ISPs and consumers

Quality of Different CDN Oracles

• CDN sites stratified by quality: Air Asia (AA) is  hosted on AkamaiEdge, ABC Streaming Video  (AB) on LimeLight Within each quality group, performance is similar

(14)

Conclusion

• Goal was to decrease ISP burden of BitTorrent traffic through reducing cross‐ISP traffic • Novel idea was to use CDNs as a predictor of  good routes • Idea experimentally validated through widely  deployed BitTorrent client plugin

References

Related documents

(related to values, norms and attitudes adverse to the rule of law that prevent formal reforms from becoming effective); (3) political economy constraints (including lack of

Supports input of Hiragana ( 平仮名 ), Katakana ( 片仮名 ) (full width and half width), Kanji ( 漢字 ) and Japanese punctuation.. The Japanese character input method is via

The green death movement challenges modern funeral practices, and what Davies (2005) identifies as the “cosmetic-casket- concrete-complex.” Green or natural burial, also referred

This study explores the hypothetical acceptability and potential utility of a reusable menstrual underwear product through examining the beliefs, behaviors, and practices

Circulating tumor cells with a putative stem cell phenotype in peripheral blood of patients with breast cancer.. Gleghorn JP, Pratt ED, Denning D, Liu H, Bander

Change Coaching Process &amp; Model, Coaching Skills, &amp; Tools Change Coaching-related Theories External Coach, Internal Coach, Peer Coach, Manager Coach, Group Coach

In this paper, we propose a framework for adaptive Hit or Miss Transform, where the HMT is defined for gray level images using the notion of adaptive structuring elements.. We present

Bio mimicry is a technological approach to engaging with the natural world which looks to nature is an intellectual source to solve human problems.. The current problems of