Jinfeng Hu

Granary: architecture of object oriented Internet storage service

Granary is a public storage service on the Internet that has two distinguished goals in comparison with previous projects. First, it is object-oriented, and thereby supports attribute-level data query. Second, it is very flexible to the system environment, i.e., it can be deployed in a grid-like environment, a peer-to-peer-like environment, or even a compromised one. In this paper, we present Granarys architecture, as well as some of its significant components that are designed in adherence to these two goals, including the node-collection protocol PeerWindow, the routing infrastructure Tourist, and the object-index management algorithm PB-link Tree. An implementation of Granary is in development and intended to be deployed in a campus scale

Achieving Reliability through Replication in a Wide-Area Network DHT Storage System

It is a challenge to design and implement a wide-area distributed hash table (DHT) which provides a storage service with high reliability. Many existing systems use replication to reach the goal of reliability. However, maintaining availability and consistency of the replicas becomes a major hurdle. A reliable storage system needs to recover lost and inconsistent replicas, but any recovery strategy will lead to extra workloads which affect the throughput of the system. This paper explores these problems and provides a possible solution. We argue that our approach not only keeps eventual consistency of replicas but also quickens the spread of updates. We use an adaptive recovery strategy to guarantee the reliability of replicas as well as bandwidth saving. With a simulation result better than epidemic algorithms, we have also implemented and deployed a DHT system using strategies mentioned in this paper, and integrated it into Granary - a storage system distributed in 20 servers in 5 cities. Granary and the DHT system have run over half a year and provide a reliable storage service to several hundred users.

PeerWindow: an efficient, heterogeneous, and autonomic node collection protocol

Nodes in peer-to-peer systems need to know the information about others to optimize neighbor selection, resource exchanging, replica placement, load balancing, query optimization, and other collaborative operations. However, how to collect this information effectively is still an open issue. In this paper, we propose a novel information collection protocol, PeerWindow, with which each node can collect a large amount of pointers to other nodes at a very low cost. Compared to existing protocols, PeerWindow is 1) efficient, the cost of collecting 1,000 pointers being less than 1 kbps in a common system environment, 2) heterogeneous, nodes with different capacities collecting different amounts of information, and 3) autonomic, nodes determining their bandwidth cost for node collection by themselves and adjusting it dynamically. PeerWindow can be used in many existing peer-to-peer systems and has tremendous potential for future expansions.

Genius: peer-to-peer location-aware gossip using network coordinates

The gossip mechanism could support reliable and scalable communication in large-scale settings. In large-scale peer-to-peer environment, however, each node could only have partial knowledge of the group membership. More seriously, because the node has no global knowledge about the underlying topology, gossip mechanism incurs much unnecessary network overhead on the Internet. In this paper, we present Genius, a novel peer-to-peer location-aware gossip. Unlike many previous location-aware techniques which utilize BGP or other router-level topology information, Genius uses the network coordinates map produced by Vivaldi as the underlying topology information. By utilizing the information, Genius could execute near-preferential gossip, that is, the node should be told the gossip message by nodes as close as possible, through which much unnecessary ‘long-range’ communication cost could be reduced. Further, the node direction information inherited in the coordinate space is exploited. We present preliminary experimental results which prove the feasibility of our scheme.

The Flexible Replication Method in an Object-Oriented Data Storage System

This paper introduces a replication method for object-oriented data storage that is highly flexible to fit different applications to improve availability. In view of semantics of different applications, this method defines three data-consistency criteria and then developers are able to select the most appropriate criteria for their programs through storage APIs. One criterion realizes a quasi-linearizability consistency, which will cause non-linearizability in a low probability but may not impair the semantic of applications. Another is a weaker one that can be used by many Internet services to provide more read throughput, and the third implements a stronger consistency to fulfill strict linearizability. In addition, they all accord with one single algorithm frame and are different from each other in a few details. Compared with conventional application-specific replication methods, this method has higher flexibility.

Twins: 2-hop Structured Overlay with High Scalability

How to build an efficient P2P overlay network on a large-scale system is still in the air. Pastry-like p2p overlays have low maintenance costs because of their log(N)-sized routing tables. However their lookup efficiency is quite low. One-hop overlays, although having high routing efficiency, can not scale to large systems because of its high maintenance cost. In this paper, we present a novel structured overlay network, Twins. Routing in Twins can be accomplished in 2 hops in very high probability. With a report-based multicast maintenance algorithm, the overlay network consumes very low maintenance cost in presence of large-scale and highly dynamic network environments. The experimental results indicate that, when the system running over a network of 5,000,000 peers, each peer consumes only 6 messages per second for maintenance, and the routing latency is only 2 hops in a very high probability of 0.99.

Gemini: Probabilistic Routing Algorithm in Structured P2P Overlay

In this paper, we propose a new structured overlay protocol, which is more efficient and scalable than previous ones. We call it Gemini, because its routing table consists of two parts, one containing nodes with common prefix and the other containing nodes with common suffix. Gemini routes messages to their destinations in just 2 hops, with a very high probability of 0.99. When running in a p2p network of 5,000,000 peers, each Gemini node consumes only 6 messages per second for maintenance. This cost, as well as routing table size, varies as a \(O {\sqrt{N}}\) function to the overlay scale, so Gemini can also run well in an even larger environment.

ImDeploy: a tool for global-scale service deployment on peer-to-peer networks

Service deployment on global-scale network is a time-consuming and complicated work if the work is really done on global-scale - not only limited to a corner of near-by sites. The network condition is varying at every moment, and the only thing that predictable is un-predictability. In this paper, we propose and implement a novel tool named ImDeploy for doing global-scale service deployment. It is adaptive to varying network condition by first clustering the target hosts by their network distance and deploying service to cluster leaders, and then deploying service inside clusters. The design and the implementation of ImDeploy also considered the scalability and user interface aspects. Through evaluation, it is proved that ImDeploy is a suitable and easy-to-use tool for global-scale service deployment.

Lookup-Ring: Building Efficient Lookups for High Dynamic Peer-to-Peer Overlays

This paper is motivated by the problem of poor searching efficiency in decentralized peer-to-peer file-sharing systems. We solve the searching problem by considering and modeling the basic trade-off between forwarding queries among peers and maintaining lookup tables in peers, so that we can utilize optimized lookup table scale to minimize bandwidth consumption, and to greatly improve the searching performance under arbitrary system parameters and resource constraints (mainly the available bandwidth). Based on the model, we design a decentralized peer-to-peer searching strategy, namely the Lookup-ring, which provides very efficient keyword searching in high dynamic peer-to-peer environments. The simulation results show that Lookup-ring can easily support a large-scale system with more than 106 participating peers at a very small cost in each peer.

Improvement of Routing Structure in P2P Overlay Networks

Although peer-to-peer overlays have been elaborately studied, they still have redundant overheads. Pastry and Chord use leaf sets to connect those nodes contiguous in numerical space, but these pointers have no contribution to efficient routing. In this paper we argue that if the mapping between a message key and its root node is determined in a smart way, leaf sets can be removed, largely decreasing system overheads. We present a novel overlay algorithm derived from Pastry, in which nodeIds have different lengths and form a Huffman set, routing tables have no empty items and root node exactly prefix-matching its keys. By these means, our approach abolishes the leaf sets with no ill effect on routing performance or other properties. Experimental results show that our approach saves about 22~25% messages cost for maintenance in comparison with Pastry in an overlay of 10,000 peers.