X hits on this document

33 views

0 shares

0 downloads

0 comments

9 / 10

4

no Oracle Oracle 100 Oracle 1000

25

no Oracle Oracle 100 Oracle 1000

6

no Oracle Oracle 100 Oracle 1000

3

Mean Leaf degree 2

1

Mean Ultrapeer degree

20

15

10

5

5

Mean Path Length 234

1

0

0

0

500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Simulation time (sec)

500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Simulation time (sec)

500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Simulation time (sec)

(a) Mean Leaf Node Degree

(b) Mean Ultrapeer Degree

(c) Mean Path Length in Overlay

Mean AS distance

3.0

2.5

2.0

1.5

1.0

0.5

no Oracle Oracle 100 Oracle 1000

Intra−AS peerings (%) LEAF

100

80

60

40

20

no Oracle Oracle 100 Oracle 1000

Intra−AS peerings (%) ULTRAPEER

100

80

60

40

20

no Oracle Oracle 100 Oracle 1000

0.0

0

0

500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Simulation time (sec)

500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Simulation time (sec)

500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Simulation time (sec)

(d) Mean AS distance in Underlay

(e) Intra-AS peerings (%) for Leaf nodes

(f) Intra-AS peerings (%) for Ultrapeers

Figure 3: Plots showing comparison of metrics in Gnutella simulations

experimentally verify that all Querys that are satisfied in unbiased Gnutella network, are also satisfied in the biased Gnutella network. We find, as predicted by the simulations, that with biased neigh- bor selection, the number of Query and QueryHit messages de- creases (60% reduction in Query, 12% reduction in QueryHit) and that the messages tend to stay within the ASes.

6.

SUMMARY AND FUTURE WORK

P2P systems build their overlay topology largely agnostic of the Internet underlay. To overcome this, we propose to use an oracle hosted by the ISPs, so that ISPs and P2P users can cooperate for im- proved performance. Such an oracle can be queried by P2P nodes while choosing neighbors or while deciding from whom to down- load content and it will rank the possible neighbors of the query- ing node according to a locality indication. We propose metrics to evaluate the effectiveness of using an oracle and show that us- ing the oracle allows the overlay topologies to maintain the graph properties like small diameter, small mean path lengths and node degree, while at the same time, tremendously increasing their net- work locality (lesser mean AS distance, larger number of intra-AS peerings). Even the ability of the network to route arbitrary traf- fic patterns with low congestion, while reduced, is still reasonable. These results along with results on improved scalability and net- work performance are obtained relying on graph based simulations, packet level simulation of an actual P2P system, as well as experi- ments with a modified P2P client in a testlab.

We are in the process of experimenting with the oracle scheme in Planetlab to increase the scale of our experiments and to test the interactions of the modified P2P clients with unmodified ones. We have realized the oracle as a Web server, which relies on a dynamic database and are in the process of installing Gnutella and Bittorrent

clients on Planetlab nodes. The Bittorrent client will consult the oracle once it gets the node list from the tracker. Alternatively the tracker may consult the oracle, to keep its list of Bittorrent nodes sorted according to distance from the querying nodes.

As more of the P2P traffic is localized within an ISP the avail- able bandwidth may increase as it is no longer capped by the peer- ing links [10]. This could lead to a usage increase which in turn may again complicate the traffic engineering problem. Yet, even this situation can be addressed by the oracle, as it can take the ISP topology and its bottlenecks into account when trying to rank the possible P2P clients.

In a next step we plan to design simple, provably good, and ex- perimentally well-behaved distributed algorithms for P2P neigh- borhood selection that take full advantage of such an oracle. We want to experiment with recent revelations of user behaviour and file sharing distributions (e.g. [38, 39]) in SSFNet, and also wish to compare the performance of oracle with latency-based protocols for neighbor selection. Computing the flow conductance of larger graphs, and exploring its relationship with lower resilience to churn is another task. An important issue that we intend to investigate is the trade-off between locality and congestion. Certainly, if strict locality is enforced (i.e., a file is always retrieved from the closest peer), there are situations where peers can encounter a high conges- tion. Hence, flexible schemes are needed that will fetch files from nearby peers if there is no congestion and otherwise will switch to more remote peers. This will eventually enable us to develop a the- oretical model to investigate the question, what is the optimal level of locality for an overlay system.

ACM SIGCOMM Computer Communication Review

39

Volume 37, Number 3, July 2007

Document info
Document views33
Page views33
Page last viewedTue Jan 24 15:50:58 UTC 2017
Pages10
Paragraphs403
Words11210

Comments