Comcast's ISP Experiences In a P4P Technical Trial

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

Comcast is a large broadband ISP, based in the U.S., serving the majority of its customers via cable modem technology. A trial was conducted in July 2008 with Pando Networks, Yale, and several ISP members of the P4P Working Group, which is part of the Distributed Computing Industry Association (DCIA). Comcast is a member of the P4P Working Group, whose mission is to work with Internet service providers (ISPs), peer to peer (P2P) companies, and technology researchers to develop "P4P" mechanisms that accelerate distribution of content and optimize utilization of ISP network resources. P4P theoretically allows P2P networks to optimize traffic within each ISP, reducing the volume of data traversing the ISP's infrastructure and creating a more manageable flow of data. P4P can also accelerate P2P downloads for end users. P4P's so-called "iTracker" technology was conceptually discussed with the IETF at the Peer to Peer Infrastructure (P2Pi) Workshop held on May 22, 2008, at the Massachusetts Institute of Technology (MIT). This work was discussed in greater detail at the 72nd meeting of the IETF, in Dublin, Ireland, in the ALTO BoF on July 29, 2008. Due to interest from the community, Comcast shared P4P trial data at the 73rd meeting of the IETF, in Minneapolis, Minnesota, in the ALTO BoF on November 18, 2008. Since that time, discussion of iTrackers and alternative technologies has continued among participants of the ALTO working group. The P4P trial was conducted, in cooperation with Pando, Yale, and three other P4P member ISPs, from July 2 to July 17, 2008. This was the first P4P trial over a cable broadband network. The trial used a Pando P2P client, and Pando distributed a special 21 MB licensed video file as in order to measure the effectiveness of P4P iTrackers. A primary objective of the trial was to measure the effects that increasing the localization of P2P swarms would have on P2P uploads, P2P downloads, and ISP networks, in comparison to normal P2P activity.

There were five different swarms for the content used in the trial. The first was a random P2P swarm, as a control group. The second, third, and fourth used different P4P iTrackers: Generic, Coarse Grained, and Fine Grained. The fifth was a proprietary Pando mechanism. (The results of the fifth swarm, while very good, are not included here since our focus is on open standards and a mechanism which may be leveraged for the benefit of the entire community of P2P clients.) Comcast deployed an iTracker server in its production network to support this trial, and configured multiple iTracker files to provide varying levels of localization to clients. In the trial itself, a P2P client begins a P2P session by querying a pTracker, which runs and manages the P2P network. The pTracker occasionally queries the iTracker, which in this case was maintained by Comcast, the ISP. Other ISPs either managed their own iTracker or used Pando or Yale to host their iTracker files. The iTracker returns network topology information to the pTracker, which then communicates with P2P clients, in order to enable P2P clients to make network-aware decisions regarding peers. The Pando client was enabled to capture extended logging, when the version of the client included support for it. The extended logging included the source and destination IP address of all P2P transfers, the number of bytes transferred, and the start and end timestamps. This information gives a precise measurement of the data flow in the network, allowing computation of data transfer volumes as well as data flow rates at each point in time. With standard logging, Pando captured the start and completion times of every download, as well as the average transfer rate observed by the client for the download. Pando served the data from an origin server external to Comcast's network. This server served about 10 copies of the file, after which all transfers (about 1 million downloads across all ISPs) were performed purely via P2P. The P2P clients in the trial start with tracker-provided peers, then use peer exchange to discover additional peers. Thus, the initial peers were provided according to P4P guidance (90% guidance based on P4P topology, and 10% random guidance), then later peers discover the entire swarm via either additional announces or peer exchange.

Trial data was collected by Pando Networks and Yale University, and raw trial results were shared with Comcast and all of the other ISPs involved in the trial. Analysis of the raw results was performed by Pando and Yale, and these organizations delivered an analysis of the P4P trial. Using the raw data, Comcast also analyzed the trial results. Furthermore, the raw trial results for Comcast were shared with Net Forecast, Inc., which performed an independent analysis of the trial for Comcast.

During the trial, downloads peaked at 24,728 per day, per swarm, or nearly 124,000 per day for all five swarms. The swarm size peaked at 11,703 peers per swarm, or nearly 57,000 peers for all five swarms. We observed a comparable number of downloads in each of the five swarms. For each swarm, below gives the number of downloaders per swarm from Comcast that finished downloading, and the number of downloaders from Comcast that canceled downloading before finishing. Characteristics of P4P Swarms: Swarm Completed Downloads Cancellations Total Attempts Cancellation Rate Random (Control) 2,719 89 2,808 3.17% --------- --------- ----------- ---------- ----------- P4P Fine Grained 2,846 64 2,910 2.20% --------- --------- ----------- ---------- ----------- P4P Generic Weight 2,775 63 2,838 2.22% --------- --------- ----------- ---------- ----------- P4P Coarse Grained 2,886 52 2,938 1.77%

The results of the trial indicated that P4P can improve the speed of downloads to P2P clients. In addition, P4P was effective in localizing P2P traffic within the Comcast network. However, we did notice that download activity in our access network increased somewhat, from 56,030 MB for Random, to 59,765 MB for P4P Generic Weight, and 60,781 MB for P4P Coarse Grained. Note that for each swarm, the number of downloaded bytes our logs report is very close to the number of downloaders multiplied by file size. But they do not exactly match due to log report errors and duplicated chunks. One factor contributing to the differences in access network download activity is that different swarms have different numbers of downloaders due to random variations during uniform random assignment of downloaders to swarms (see ). One interesting observation is that Random has higher cancellation rate (3.17%) than that of the guided swarms (1.77% to 2.22%). Whether guided swarms achieve lower cancellation rate is an interesting issue for future investigation. Impact of P4P on Downloads: Swarm Global Avg bps Change Comcast Avg bps Change Random (Control) 144,045 bps n/a 254,671 bps n/a ---------- ---------- ---------- ---------- ---------- P4P Fine Grained 162,344 bps +13% 402,043 bps +57% ---------- ---------- ---------- ---------- ---------- P4P Generic Weight 163,205 bps +13% 463,782 bps +82% ---------- ---------- ---------- ---------- ---------- P4P Coarse Grained 166,273 bps +15% 471,218 bps +85%

An analysis of the effects of P4P on upstream utilization and Internet transit was also interesting. It did not appear that P4P significantly increased upstream utilization in the Comcast access network; in essence uploading was already occurring no matter what and P4P in and of itself did not appear to materially increase uploading for this specific, licensed content. (P4P is not intended as a solution for the potential of network congestion to occur.) Random was 143,236 MB and P4P Generic Weight was 143,143 MB, while P4P Coarse Grained was 139,669 MB. We also observed that P4P reduced outgoing Internet traffic by an average of 34% at peering points. Random was 134,219 MB and P4P Generic Weight was 91,979 MB, while P4P Coarse Grained was 86,652 MB. In terms of downstream utilization, we observed that P4P reduced incoming Internet traffic by an average of 80% at peering points. Random was 47,013 MB, P4P Generic Weight was 8,610 MB, and P4P Coarse Grained was 7,764 MB. However, we did notice that download activity in the Comcast access network increased somewhat, from 56,030 MB for Random, to 59,765 MB for P4P Generic Weight, and 60,781 MB for P4P Coarse Grained. Note that for each swarm, the number of downloaded bytes according to logging reports is very close to the number of downloaders multiplied by file size. But they do not exactly match due to log report errors and duplicated chunks. One factor contributing to the differences in access network download activity is that different swarms have different numbers of downloaders, due to random variations during uniform random assignment of downloaders to swarms (see Table 1). One interesting observation is that Random has higher cancellation rate (3.17%) than that of the guided swarms (1.77%-2.22%). Whether guided swarms achieve lower cancellation rate is an interesting issue for future research.

Given the size of the Comcast network, it was felt that in order to truly evaluate the iTracker application we would need to test various network topologies that reflected its network and would help gauge the level of effort and design requirements necessary to get correct statistical data out of the trial. In all cases, iTrackers were configured with automation in mind, so that any successful iTracker configuration would be automatically updating, rather than manually configured on an on-going basis. All iTrackers were hosted on the same small server, and it appeared to be relatively easy and inexpensive to scale up an iTracker infrastructure should P4P-like mechanisms become standardized and widely adopted.

The Fine Grain topology was the first and most complex iTracker that we built for this trial. It was a detailed mapping of Comcast backbone-connected network Autonomous System Numbers (ASN) to IP Aggregates which were weighted based on priority and distance from each other. Included in this design was a prioritization of all Peer and Internet transit connected ASNs to the Comcast backbone to ensure that P4P traffic would prefer settlement free and lower cost networks first, and then more expensive transit links. This attempted to optimize and lower transit costs associated with this traffic. We then took the additional step of detailing each ASN and IP aggregate into IP subnets down to Optical Transport Nodes (OTN) where all Cable Modem Termination Systems (CMTS) reside. This design gave a highly localized and detailed description of the Comcast network for the iTracker to disseminate. This design defined 1,182 iTracker node identifiers, and resulted in a 107,357 line configuration file. This iTracker was obviously the most time-consuming to create and the most complex to maintain. Trial results indicated that this level of localization was too high, and was less effective compared to lower levels of localization.

Given the level of detail in the Fine Grain design, it was important that we also enable a high-level design which still used priority and weighting mechanisms for the Comcast backbone and transit links. The Coarse Grain design was a limited or summarized version of the Fine Grain design, which used the ASN to IP Aggregate and weighted data for transit links, but removed all additional localization data. This insured we would get similar data sets from the Fine Grain design, but without the more detailed localization of each of the networks off of the Comcast backbone. This design defined 22 iTracker node identifiers, and resulted in a 998 line configuration file. From an overall cost, complexity, risk, and effectiveness standpoint, this was judged to be the optimal iTracker for Comcast. Importantly, this did not require revealing the complex, internal network topology that the Fine Grain did. Updates to this iTracker were also far simpler to automate, which will better ensure that it is accurate over time, and keeps administrative overhead relatively low. However, the differences, costs, and benefits of Coarse Grain and Generic Weighted (see below) likely merit further study.

The Generic Weighted design was a copy of the Coarse Grained design but instead of using ISP-designated priority and weights, all weights were defaulted to pre-determined parameters that the Yale team had designed. All other data was replicated from the Coarse Grain design. Providing the information necessary to support the Generic Weighted iTracker was roughly the same as for Coarse Grain.

Raw data is presented in this document. We did not normalize traffic volume data (e.g. upload and download) by the number of downloads in order to preserve this underlying raw data. We also recommend that readers not focus too much on the absolute numbers, such as bytes downloaded from internal sources and bytes downloaded from external sources. Instead, we recommend readers focus on ratios such as the percentage of bytes downloaded that came from internal sources in each swarm. As a result, the small random variation between number of downloads of each swarm does not distract readers from important metrics like shifting traffic from external to internal sources, among other things. We also wish to note that the data was collected from a sample of the total swarm. Specifically, there were some peers running older versions of the Pando client that did not implement the extended transfer logging. For those nodes, which participated in the swarms but did not report their data transfers, we have download counts. The result of this is that, for example, the download counts generated from the standard logging are a bit higher than the download counts generated by the extended logging. That being said, over 90% of downloads were by peers running the newer software, which we believe shows that the transfer records are highly representative of the total data flow. In terms of which analysis was performed from the standard logging compared to extended logging, all of the data flow analysis was performed using the extended logging. Pando's download counts and performance numbers were generated via standard logging (i.e. all peers report download complete/cancel, data volumes, and measured download speed on the client). Yale's download counts and performance numbers were derived via extended logging (e.g. by summing the transfer records, counting IP addresses reported, etc.). One benefit of having two data sources is that we can compare the two. In this case, the two approaches both reported comparable impacts.

One objective of this document is to share with the IETF community the results of one P4P trial in a large broadband network, given skepticism regarding the benefits to P2P users as well as to ISPs. From the perspective of P2P users, P4P potentially delivers faster P2P downloads. At the same time, ISPs can increase the localization of swarms, enabling them to reduce bytes flowing over transit points, while also delivering an optimized P2P experience to customers. However, an internal analysis of varying levels of iTracker adoption by ISPs leads us to believe that, while P4P-type mechanisms are valuable on a single ISP basis, the value of P4P increases dramatically as many ISPs choose to deploy it. We believe these results can inform the technical discussion in the IETF over how to use iTracker mechanisms. Should such a mechanism be standardized, the use of ISP-provided iTrackers should probably be an opt-in feature for P2P users, or at least a feature of which they are explicitly aware of and which has been enabled by default in a particular P2P client. In this way, P2P users could choose to opt-in either explicitly or by their choice of P2P client in order to choose to use the iTracker to improve performance, which benefits both the user and the ISP at the same time. Importantly in terms of privacy, the iTracker makes available only network topology information, and would not in its current form enable an ISP, via the iTracker, to determine what P2P clients were downloading what content. It is also possible that an iTracker type of mechanism, in combination with a P2P cache, could further improve P2P download performance, which merits further study. In addition, this was a limited trial that, while very promising, indicates a need for additional technical investigation and trial work. Such follow-up study should explore the effects of P4P when more P2P client software variants are involved, with larger swarms, and with additional and more technically diverse content (file size, file type, duration of content, etc.).

There are no security considerations to include at this time.

There are no IANA considerations in this document.

The authors wish to acknowledge the hard work of all of the P4P working group members, and specifically the focused efforts of the teams at both Pando and Yale for the trial itself. Finally, the authors recognize and appreciate Peter Sevcik and John Bartlett, of NetForecast, Inc., for their valued independent analysis of the trial results.