A Framework for Peer Selection Analysis
As previously stated, our aim is to develop a rigorous framework to unveil the “networkawareness” exhibited by P2P-TV applications, i.e., which network parameters current P2P-TV systems take into account when distributing the stream. We define a flexible framework that allows us not only to inspect the level of “awareness” of a P2P system with respect to the underlying network, but also to assess whether peers behave fairly with respect one to another, i.e. if the peers are incentivized to the mutual data exchange. In particular, we consider:
• AS(p): the Autonomous System where peer p is located.
• CC(p): the Country, which peer p belongs to.
• NET(p): the subnetwork, which peer p belongs to.
• HOP(p, e): the IP hop-distance between peers p and e.
• SYM(p, e): the symmetry of byte-wise data exchanges between peers p and e.
Let p ∈ W denote a peer that belongs to the NAPA-WINE set W. Let P(p) denote the set of contributing peers, p exchanges data with. That is, P(p) is composed by the peers to which p transmitted or/and from which p received some video information. Let U(p) denote the subset of peers to which p is uploading video content, and D(p) the subset from which p is downloading video from. U(p) and D(p) are two (non disjoint) subsets of P(p), and U(p) ∪ D(p) = P(p). Let e ∈ P(p) be an arbitrary peer that exchanges traffic with p. Denote by B(p, e) the amount of bytes transmitted from p to e, so that B(e, p) represents the amount of bytes received by p from e.
Consider now a generic network parameter X(·), and denote with X(p, e) ∈ X the observed value ofX(·) for the pair (p, e). We partition P(p) into two classes based onX(p, e), such that one class should intuitively be preferred from the application (e.g., good vs bad peers). More formally, we partition the support X into two disjoint sets: the preferred set XP and its complement XP , such that XP ∪ XP = X and XP ∩ XP = ∅.
For the ease of notation, let 1P (p, e) be the identity function which takes the value of 1 if X(p, e) ∈ XP and 0 otherwise; similarly, 1P (p, e) = 1 − 1P (p, e). Without loss of generality, let us focus on the upload traffic of a NAPA-WINE peer p ∈ W, and let us define: PeerU|P (p) = X e∈U(p) 1P (p, e).
Preliminary Analysis and Issues
Given the black box approach based on passive measurement, several issues could undermine the significance of the results unless carefully dealt with. The first issue is that the NAPA-WINE peers induced a bias during the experiments. Recall that among NAPA-WINE peers there are several high-bandwidth peers, located in Europe only. Furthermore, all peers within the same institution are in the same LAN, and AS. This possibly represents an uncommon population subset.
A quantification of the induced bias is given in Tab. 3.2. It reports the percentage of i) NAPAWINE peers over all peers observed during each experiment, and ii) bytes exchanged among NAPA-WINE peers over all exchanged bytes. Results are reported considering contributors only, or all peers. As first important remark, NAPA-WINE peers clearly prefer to exchange data among them. For example, considering contributors in the PPLive experiment, NAPA-WINE peers contribute to more than 3.5%of exchanged data, even if they represent less than 1%of the contributing peers. Similarly, they are about 10% and 30% of peers for SopCast and TVAnts respectively, but they contribute to 18% and 56% of exchanged bytes. We stress that by restricting the analysis to the set of peers other than NAPA-WINE, it will be possible to highlight and quantify which properties of the NAPA-WINE peers cause such a strong bias. To solve the issue concerning the induced bias, we introduce the set P′(p) ⊂ P(p). Subset P′(p) is constituted by the peers in P(p) excluding the NAPA-WINE peers, formally P′(p) = P(p) \W. We evaluate the preference metrics also over the filtered set, getting P′ D, P′ U ,B′ D ,B′ U , accordingly. Intuitively, restricting the observation to P′ is equivalent to consider peers not involved in the experiment. For example, we expect that a preference versus a metric noticed in the full contributor set should be noticeable also in the set deprived of NAPA-WINE peers. In case the bias is still evident, this means that the preference was not artificially induced by NAPA-WINE peers.
AS and Country Awareness
We first turn our attention to location awareness by considering the AS and CC metrics. Considering download direction, it can be seen that SopCast is unaware of AS location. Indeed, PD is almost equal to BD, which suggests that peers in the same AS are not preferentially selected to download data from. On the contrary, both PPLive and TVAnts show higher AS-awareness. Considering non-NAPA-WINE contributors, a PPLive peer downloads from P′ D=0.6% of peers B′ D =6.5% of traffic, i.e., there is a byte preference 10 times larger than a peer preference. The same factor holds including NAPA-WINE peers (which then do not bias the results). Similarly, for TVAnts, B′ D =7.6% of the bytes are downloaded from P′ D =3.3% of the non-NAPA-WINE contributors, i.e., a B′ D /P′ D ratio equal to 2. Also, notice that 0.04% of all peers are in the same AS of NAPA-WINE peers in case of PPLive, and 3.6% in case of TVAnts. Still, as 1.3% of the contributing peers are located in the same AS for PPLive, and 13.5% for TVAnts, we can conclude that PPLive exhibits a stronger preference for peers within the same AS than TVAnts.
Looking at the downloaded traffic with respect to the peer Country, we notice that almost the same percentages are observed as in the AS preference case. Since two peers in the same AS are also located within the same Country, we can conclude that no country preference is shown, i.e., the CC preference is due to the AS preference. Finally, considering the upload directions, similar conclusions can be drawn.
We now evaluate the potential preference to exchange traffic with peers in the same subnet (NET). The set of peers in the same subnet includes only NAPA-WINE peers, i.e., P′ = ∅. Results show that also in this case, PPLive and TVAnts only exhibit NET awareness, for both upload and download directions. Indeed, about 10% and 18% of the bytes are received from about 1% and 7% of hosts which are in the same subnet respectively. Conversely, SopCast does not show any evidence of subnet awareness. However, the NET preference can be also enforced by the AS preference. Looking at the ratio between P over B for the AS and NET preferences, we observe that they are very similar. This points out that peers in the same autonomous system but not in the same NET are equally preferred as the peers in the same NET (and in the same AS). Therefore, the AS preference is stronger than the NET preference. Notice also that the AS locality is overall quite marginal, so that the majority of the traffic is still coming from other ASs. As such, there is large margin to improve the network friendliness of P2P-TV applications.
We also investigate the HOP count preference. In this case, no particular evidence of preference toward shorter paths is underlined. Indeed, looking at the non-NAPA-WINE peers, almost no difference emerges comparing P′ and B′. Only TVAnts shows a small preference to download from closer nodes.
To further testify this finding, Fig. 3.2 reports the Cumulative Distribution Function (CDF) of contacted peers (solid line) and of the received bytes (dashed line) versus the distance between peers in hop count, not including the NAPA-WINE peers. TVAnts only shows a slight commitment to the closest peers, while SopCast and PPLive seem to ignore peer distance considering the hop number.
Table of contents :
1.2 The Big Picture
I Measuring network awareness
2 Preliminary discussion
2.1 Related Work
2.2 Passive Data Set
3 Passive Analysis
3.1 Preliminary Results
3.2 A Framework for Peer Selection Analysis
3.3 Experimental Results
3.4 Dynamics of Contacted Peers
4 Hybrid Analysis
4.2 Experimental Results: Path-wise Metric
4.3 Experimental Results: Peer-wise metric
5 A comprehensive framework to test Network Awareness
5.1 Analysis Process
5.2 Features Definition
5.3 Metric Definition
5.4 Experimental Results
II Implementing network awareness
6 Simulation Analysis
6.1 Related work
6.2 Framework Description
6.3 Simulation Results: Impact of L7 and L3
6.4 Simulation Results: L3/L7 Interaction
7 Emulation Analysis
7.1 Related work
7.2 ModelNet-TE Emulator
7.3 Scenario and methodology
7.4 Experimental Results
8.2 Future Work
A List of publications
A.2 Under Review