Arun VenkataramaniDonald Towsley

Presented by: Shiqi Chen, Ionut Trestian

Late 90s client-server architectures dominated data transfers on Internet

Alternatives emerged:◦ Unstructured p2p file sharing◦ Content Distribution Networks◦ Structured Distributed Hash Tables◦ Publish subscribe◦ End-System Multicast

Main point: make data location Independent

Problems with existing p2p systems◦ Free riding◦ Central point of failure (BitTorrent)

Simple and robust incentive strategy (tit-for-tat) is a deterrent for free riding.

Can swarms form the basis of a universal architecture for the Internet and if so what is an appropriate architecture?

Fundamental difference◦ Uswarm (one huge swarm for all data transfers)◦ BitTorrent (one swarm per file)

Typical transfer multipoint-to-point

Data is location independent and self-verifying

Designed to send bulk data but can send dynamic data too.

Intent Resolution Service – translating intent (what a peer is looking for, URL, RSS feed) to metadata

IRS, Modern search engine, web server that serves metadata

Metadata Resolution Service returns set of peer addresses (tracker in BitTorrent), distributed here.

Peers translate metadata to addresses(3 ways)◦ Logically centralized tracking service similar to DNS◦ In-network support for tracking (gateway router

intercepts and processes resolution requests)◦ Peer-to-Peer tracking

Centralized trackers suited for pull data transfer but not fur push data transfer.

Enables location based caching

Peers have roles depending on their economic relationship with their peers and network service provider.

Some peers always ready to send data, social network based trust possible.

Live streaming – p2p live streaming Semi-autonomous peer system – push

contents to set-top boxes Human-centric applications – facilitates

push based applications

Fundamental benefits of uswarm over isolated swarms

◦ Post-popularity – BitTorrent robust to flash crowds but treats bad unpopular content.

◦ Block-availability - Allow peers to change blocks across different content.

◦ Robust tracking – BitTorrent is inherently unscalable. Fractured content

Download capacity O(log n) greater

Also more robust to failures

BitTorrent – central server (single point of failure)

Low tracker availability is a significant problem for users.

Solutions◦ Replicated trackers◦ Integration of DHTs with BitTorrent clients.

Our approach, a combination of:◦ Massively replicated tracking.

Similar to DNS (hierarchical)

◦ Peer-to-peer gossip. Controlled flooding Active gossip

◦ In-network tracking. Couples a uswarm tracker with in-network caches.

Traffic engineering

Today’s Internet constraints the routing choice of both end users and ISPs

Bring ISPs into the picture (tracker router): return a set of peers that result in the most load-balanced traffic assignment.

More complex scenarios required.

Which replicated tracking mechanism is best suited.

Implicitly assumed that download rate is the performance metric. For some applications this is not the case.

In-network caching raises the question of placement and replacement strategies.

Incentives in isolated swarmsBitTorrent is not robust to selfish peer behaviorBittyrant: selfish algorithm benefits every peer

irrespective of how many peers are using it.

Open questions:Estimating peer upload and download

capacitiesAnalyzing if the proposed mechanism is

strategy proofRobustness to byzantine faulty peers

Data plane: extend Bittyrant algorithm – select peers so as to maximize download on available upload rate

Users may upload blocks from different files to download a particular file or simultaneously download several files with different utility.

Sort peers according to dp/up weighted by utility

Control planetit-for-tat: keep track of the number of

metadata requests a neighbor helps resolve. Peers choose to help neighbors that have been most useful in the past.

Dynamic topology adaptation: peers prefer other peers with similar content interest for neighbors, resembling a semantic social network.

How to measure the performance of control plane?

Peer selection in the data plane at fine time scales.

Peer selection in the control plane that operates over slower time scales

Movement of peers across interest clusters based on their content access pattern.

Opening a connection has cost.

So S will not benefit even he increase the number of connections if we charge a fixed cost per connection.

tit-for-tat: Pairwise Nash equilibrium exists but the loss of efficiency is unbounded.

If considering routing topology and congestion control, then Nash equilibrium only exists for simple topologies and the loss of efficiency can be arbitrarily large.

Goal: Interactive strategic peer selection while taking topology and allocation cost into account.

Pricing: upload bandwidth may not be free or even linearly priced.

Indirect trading: agents with low upload capacity may appoint other agents to participate in uswarm on its behalf. P2P detour routing may be implemented as a service on top of uswarm.

Workload and network conditions: Block availability, TCP’s inefficiency, block-based tit-for-tat strategy, noisy information on the efficiency or vulnerability.

DoI-resistant information dissemination service – if a user seeks a file at least one copy of which exists in the network then she could be able to retrieve it unless the underlying network is physically partitioned by faulty or malicious nodes.

Basic idea: to use massive replication to ensure that a peer can reach at least one replica.Knowledge of routing topology: select neighborsReplicating content: DoS attacks do not make

information unavailable.

Instrument a small number of Bittyrant clients to passively monitor transfer performance and contribute measurement.

It is incentive-compatible for peers to contribute measurements because peers with more information about the capacity of other peers can make better strategic choices.

uswarm is tolerant to delays or interrupted point-to-point connections by design.

RAPID (Resource Allocation Protocol for Intentional DTN routing): enables sophisticated knobs to intentionally optimize performance metrics of interest.

Push-based application: advertising and trading of content according to individual interest.

Model single-swarm and multi-swarm behavior accounting for strategic peer behavior.

Topology adaptation model and implementation of DoI-resistant dissemination service.

Implement a modified socket API to enable application use. Investigate interaction with traffic engineering and virtualized architectures. Test for security.

Enable a robust dissemination service resistant to denial of information attacks.

Foster the growth of novel human-centric applications.

Enable delay-tolerant data transfer for poorly connected environments.

Thank you !Questions ?