“ P4P is a flexible framework that can be used to enable Internet service providers (ISPs) and peer-to-peer (P2P) software distributors to work jointly and cooperatively. ” -- Haiyong Xie

P4P Field Tests

P4PWG has been conducting two sets of P4P field tests as of March, 2008:

1. Yale University, Pando Networks and Verizon have conduct P4P field tests and released the test results in the past month.
2. Yale University, Pando Networks and Telefonica have been conducting P4P field tests and will release the test results soon.

Summary of Field Test Results

We have observed that both ISP and P2P can benefit from P4P, based on the results of our field tests on Verizon network. In particular, end users benefit from a dramatic improvement in data delivery speed, which results in faster downloads for users; and ISPs saved by shifting traffic from external to internal links, and by routing internally much more efficiently.

P2P Perspective

• The improvement of download rates for data transfers among FTTH clients was 205% on average.
• The improvement of download rates for FTTH clients downloading from all clients was 23%.

ISP Perspective

• P4P yields a dramatic drop in data delivery average hop count (from 5.5 to 0.89), which equates to lower cost to ISPs.
• Total external (peering) link load dropped by as much as 42% (outbound) and 35% (inbound).
• Normalized load on internal backbone links dropped by as much as 71% on average.

Configuration of Field Tests

P4P Components

In both field tests, Haiyong Xie (Yale University) run the prototyped iTracker, Pando Networks run the appTracker and distribute a 20MB file (Pando users run the Pando client software to download the file), and carriers (Telefonica and Verizon) provides its network map information.

In each test, Pando Networks sets up two parallel swarms (of approximately the same size and distribution), one called "Native P2P" (vanilla Pando P2P), the other called "P4P" (P4P-enabled Pando P2P).

Information Flow

iTracker processes the network map information (provided by Telefonica and Verizon) and P2P network status information (provided by Pando appTracker).

appTracker communicates with iTracker through query-response interfaces. appTracker frequently updates iTracker with its current P2P network status through queries, and iTracker responds to appTracker's queries with P4P-specific guidelines/suggestions dynamically computed on the fly.

Logs are collected for data analysis. Only those clients in Telefonica and Verizon's networks are P4P-enabled.

Comments on Test Configurations

Many ISPs implement traffic shaping policies on interdomain links to throttle P2P transfer rates. The current P2P improvement results do not reflect this factor. We anticipate that P4P may further improve P2P performance in this setting.

In the field test, we did not include interdomain policy into consideration yet; but we will do so in future tests.

Detailed Results of Field Test

Transit Traffic

Overall Traffic Distribution

• Download traffic between external clients (outside Verizon) is approximately the same for Native and P4P
• The amounts of traffic flowing into Verizon and out of Verizon are much higher without P4P
• P4P increases traffic localized to Verizon clients to be approximately 6.6 times.
• The total amount of traffic is approximately the same for Native and P4P.

Distribution of Traffic within Verizon Network

P4P effectively reduces the amount of cross-metro-area traffic within Verizon network (not surprisingly). Specifically, in Native test, the amount of intra-metro-area traffic is only 6.27% (of all traffic inside Verizon); while in P4P test, this percentage increases to 57.98%.

Distance of Data Delivery

Consider the data transfers among Verizon clients only. Note that if Verizon's internal routing and BGP routing table were available, we should also be able to find the entry/exit points for each external client and thus compute the number of hops that the inbound and outbound traffic traverses.

We use a metric referred to as BDP (Weighted Average Bit-Distance Product) to measure the distance of data delivery. BDP = sum of distance (metro hop count) each bit traverses / total number of bits.

P4P results in approximately 5 times reduction (0.89 vs 5.5) in BDP. Note that the average number of metro hops between two metros is 6.2, and P2P results in BDP close to 6.2 due to its random peering (skewed a bit due to not all data being delivered from random peers)

Normalized Link Load

Consider the data transfers among Verizon clients only. Noting that P4P results in approximately 3 times more cross-metro-area traffic, we normalize the average traffic loads on links in order to make fair comparisons.

P4P results in 71% lower load on backbone links on average. Specifically, Native results in an average load of 71,666,279 bytes, while P4P results in an average normalized load of 25,071,685 bytes; in other words, Native results in 1.85 times higher (normalized) load than P4P.

Download Completion Time

Consider the metric of average session download rate. For each client and all log lines which recorded the events that the client downloaded from others, we compute the average rate (i.e., total amount of traffic divided by the difference between the start time of the client's first download and the end time of the client's last download). Note that this metric is different from instantaneous download rate. <-->

Consider the download completion time of Verizon clients.

P4P improves average completion time by 23%. Most of P4P clients (in Verizon network) complete downloading the data from the other clients in Verizon network. However, none of Native clients (in Verizon network) complete downloading the data only from others in Verizon.

Note that the last mile has the greater influence on the completion time and download rates. Therefore, the dramatic improvement in BDP does not result in as much improvement in the completion time and download rate.

Statistics for FTTH/FTTP Clients

Clients with FTTH/FTTP (fiber to the home / fiber to the premises) access links in Verizon network see more performance improvement.

FTTH/FTTP Clients Downloading from All Clients

P4P improves the download rates for FTTP clients by 32% (P4P results in an average download rate of 148,266, while Native results in 111,968. Unit: bytes/second).

P4P improves the average completion time for FTTP clients by 68%.

Rate of File Downloading among FTTH/FTTP Clients

P4P improves the download rates for FTTP clients (downloading from other FTTP clients) by 205%.

None of Native FTTP clients completes downloading the data solely from other FTTP clients in Verizon. However, most of the P4P FTTP clients complete downloading the data only from other FTTP clients. Thus the following figure shows the completion time for P4P tests only.

Swarm Dynamics

The swarm size is dynamic. Both swarms (Native and P4P) have approximately the same size. The swarm reached its largest size in the first 3 days, and decreased to a low number of clients afterwards.

This figure shows the dynamics of the two swarms (including all clients).

This figure shows the dynamics of the two swarms (including Verizon clients only).