|
|
|
Peer-to-Peer Systems and Networking
Overview
Our current research activities on peer-to-peer systems and networking expand from fundamental platform research to more application oriented overlay optimization technologies and a few novel applications. More specifically, we target to develop an open, non-application-specific overlay platform, including a set of common APIs, for P2P research community to rapidly prototype and, hopefully, to develop real products with much reduced effort. We study new technologies that can effectively relieve the severe tension between the P2P applications and the ISPs. We design new network coordinate systems that can pinpoint any Internet-connected computing device with an unprecedented precision. We propose high efficient search technologies based on distributed segment tree and link prediction algorithms. We build novel applications such as Info-Brokerage - a distributed, peer-to-peer Q&A application that also features social computing, SigFeed - a unified architecture for prompt event notification like RSS feed and also efficient podcast downloading, and mTreebone - a hybrid tree/mesho overlay for live video multicast, to name a few.
We also actively research towards new networking paradigms. While
there are lots of works on the incentive mechanisms, e.g.
tit-for-tat or based on social networking, for P2P applications, we
envision a virtual online market will be formed sooner or later and
virtual currency will be adopted for users to trade resources,
taking analogy of "points" or "medal" in current networked games. We established the PeerPool project,
of which the objective is to allow users to deposit, rent and loan their computing resources (CPU, bandwidth, storage, location, etc)
into or from a virtual market. To the virtual market operator, it's a bank model and to users, it's
a buy-to-let investment model.
People buy computers and rent them when not using to earn some
"money" back. Furthermore, we also envision the shift from
application-specific P2P systems to
service-oriented P2P services (SoP2P). Current P2P systems are all
application-specific, albeit file downloading or live streaming.
Benefits of SoP2P include simplification of the application
development process and progressive yet transparent improvement of
many services through the improvement of underlying p2p
technologies. It also reduces the cost for creating new services or
extending existing services. Project Members
Projects
Existing popular P2P applications range from P2P streaming (both live and on demand), P2P file swarming, P2P-CDN, and distributed storage etc. Typically, each application builds its own dedicated overlay with resources from its own participants. We term such P2P systems application-specific P2P systems (AsP2P). Existing AsP2Ps improve QoS as compared with their non-P2P counterparts, but may not meet desired QoS level in many applications yet. There are many research activities that try to improve performance of certain popular P2P applications. Once we have means to meet desired level of QoS, we can anticipate a shift from AsP2P systems to service-oriented P2P services (SoP2P). For instance, we can provide an generic data delivery service that acts as a common interface between an application and the lower-level network layers. Benefits of SoP2P include simplification of the application development process and progressive yet transparent improvement of many services through the improvement of underlying p2p technologies. It also reduces the cost for creating new services or extending existing services. We envision that future applications will be mostly composed from various services.
Generally speaking, P2P is favored more by end users than Internet Services Providers (ISPs). Currently, the traffic generated by P2P systems accounts for a major fraction of the Internet traffic, outstripping every other communication and distribution protocols and is bound to increase. Worse even, in excess of 92% P2P traffic crosses transit/peering links among ISPs. This inter-AS traffic is especially important to ISPs, as it typically affects their bottom line. P2P affects QoS levels for all subscribers of ISPs, including P2P applications themselves, because P2P protocols usually aggressively consume any available bandwidth capacity. Due to symmetrical nature of p2p system, on average 80% of upstream capacity is consumed by P2P. Given the tension incurred by P2P, there are generally two type of measures, destructive or constructive, that ISPs can take. The destructive way is to apply traffic engineering against P2P traffic like filtering or rate limiting the P2P traffic. However, such measures can end up backfiring on ISPs because users may simply take their business elsewhere. The constructive way is to handle P2P traffic more intelligently, such as provide P2P caching. However, this implies huge monetary investment. In this project, we propose an HTTP based Peer-to-Peer (HPTP) framework to relieve the tension, which is a pleasing proposal to both ISPs and end users (including both P2P users and normal web users), i.e., may lead to a win-win situation. The key idea is to HTTPify the P2P traffic so that they can be cached by the already widely deployed (web) caching proxies by both ISPs and organizations, which achieves similar effect as deploying dedicated P2P caches but at no extra cost. By HTTPifying, we mean to segment (if necessary) the files or streams into chunks, use HTTP protocol for the transport of the resultant chunks and ensure such HTTP wrapped chunks are cacheable. We present necessary tools like the HTTPifying tool and the cache detection and usability test tool HPing. We perform case study on building a cache aware delivery tree for streaming scenarios.
Network coordinates (NC) system is an efficient mechanism to predict latency between any two Internet nodes by directly calculation using the nodes' coordinates instead of explicit measurement to them. NC system can significantly reduce the active probing overhead and is therefore quite beneficial to large scale Internet applications such as peer-to-peer content distribution, multi-user gaming, etc. NC system has drawn lot of research attention, and many approaches have been proposed in the literature that can be generally classified into two categories, namely landmark-based schemes such as GNP or PIC and simulation-based systems like Vivaldi . Landmark-based schemes have the merit of faster convergence and improved accuracy, but require special management of deployed landmark servers. Partially due to this requirement, such service has not yet emerged in practice so far. Simulation-based schemes do not rely on the landmark and determine the coordinates by simulating the system with a physical model. But in their procedures, all nodes are taken into account which could degrade its consistency, robustness, and the convergence speed, due to the node dynamics. To design a practical, scalable and robust NC system, we propose a novel NC scheme, Whereis, that blends the merits of landmark-based and simulation-based schemes while removing the dependency on the priori knowledge or deployment of a landmark infrastructure. Most prominently, Whereis can position any Internet node without its explicit participance, which is not possible with any other existing NC systems, and is therefore more applicable. Moreover, by leveraging existing stable and powerful servers as passive landmarks, Whereis achieves excellent consistence and robustness and imposes less security risks.
While most P2P works focusing on one or few specific applications, we look at the problem from a more fundamental aspect. We target to develop an open, non-application-specific overlay platform, including a set of common APIs, for P2P research community to rapidly prototype and, hopefully, to develop real products with much reduced effort. One of the design goals is to allow different P2P overlays built on our platform to have a substrate over which segmented overlays can be unified into a bigger one. This will allow some peers to help others who are possibly with different tasks in their free time, under certain incentive mechanisms or based on the free marketing model. For instance, BitTorrent users may help PPLive users when they are not downloading any content or have spare bandwidth.
The vision of PeerPool is to provide infrastructure service to other popular Internet services that are resource demanding and distributed in nature. The key idea is build a common logical infrastructure via network management to provide a resource banking system through which some end users can deposit their spare resources while others can loan to improve their own business. The prominent point of PeerPool is that it aggregates peers' resources across different P2P applications, and in return make rich resources available to many P2P applications that hope to improve their QoS, through marketing peers' resources. The key features of PeerPool are: In fact, PeerPool concept is specifically motivated by CDN service. The CDN market is enabled due to high demand for data delivery with certain QoS improvement (not guarantee). CDN vendor deploys many edge servers at their own cost, but they can make money because they make good, if not full, use of their deployed resources. As long as the utilization ratio of their deployed server reaches a certain level, their incomings would exceed their expense and they earn money. In the P2P world, we realize that there are actually abundant resources that have been invested by end users while many p2p applications still suffer performance due to not having access to enough resources. So, a natural idea is to unite the resources from providing users and offer the aggregated resources to consuming users. As a general economic principle: trade makes everyone better off. With the promising of earning money with their resources, end users will have incentive to join the PeerPool (which of course depending on the rewarding rate). While recognizing that QoS of peer-to-peer applications can be guaranteed if enough resources are available or at least certainly improved if provided more resources, we want to study the following aspects: Publications
Wireless and Networking Group | Microsoft Research Asia |
