Slicing Distributed Systems (original) (raw)

Fast distributed slicing without requiring uniformity

2008

Peer-to-peer (P2P) systems have become popular as a means of sharing resources efficiently within large sets of end-user platforms. Among their many applications are file sharing , storage and Voice-over-IP (VoIP) . Ideally, such systems should be able to perform well even when the peers are heterogeneous with respect to performance-critical attributes such as bandwidth or storage capacity.

Distributed slicing in dynamic systems

… , 2007. ICDCS'07. …, 2007

Peer to peer (P2P) systems are moving from application specific architectures to a generic service oriented design philosophy. This raises interesting problems in connection with providing useful P2P middleware services capable of dealing with resource assignment and management in a large-scale, heterogeneous and unreliable environment. The slicing service, has been proposed to allow for an automatic partitioning of P2P networks into groups (slices) that represent a controllable amount of some resource and that are also ...

Divide et impera: Partitioning unstructured peer-to-peer systems to improve resource location

Achievements in European Research on Grid Systems, 2008

Unstructured P2P systems exhibit a great deal of robustness and self-healing at the cost of reduced scalability. Resource location is performed using a broadcast-like process called flooding. The work presented in this paper comprises an effort to reduce the overwhelming volume of traffic generated by flooding, thus increasing the scalability of unstructured P2P systems. Using a simple hash-based content categorization method the Ultrapeer overlay network is partitioned into a relatively small number of distinct subnetworks. By employing ...

Effective load-balancing of peer-to-peer systems

The growing popularity of peer-to-peer (P2P) systems has necessitated the need for managing huge volumes of data efficiently to ensure acceptable user response times. Dynamically changing popularities of data items and skewed user query patterns in P2P systems may cause some of the peers to become bottlenecks, thereby resulting in severe load imbalance and consequently increased user response times. An effective load-balancing mechanism becomes a necessity in such cases. Such load-balancing can be achieved by efficient online data migration/replication. While much work has been done to harness the huge computing resources of P2P systems for high-performance computing and scientific applications, issues concerning load-balancing with a view towards faster access to data for normal users have not received adequate attention. Notably, the sheer size of P2P networks and the inherent dynamism of the environment pose significant challenges to load-balancing. The main contributions of our proposal are three-fold. First, we view a P2P system as comprising clusters of peers and present techniques for both intra-cluster and inter-cluster load-balancing. Second, we analyze the trade-offs between the options of migration and replication and formulate a strategy based on which the system decides at run-time which option to use. Third, we propose an effective strategy aimed towards automatic self-evolving clusters of peers. Our performance evaluation demonstrates that our proposed technique for inter-cluster load-balancing is indeed effective in improving the system performance significantly. To our knowledge, this work is one of the earliest attempts at addressing load-balancing via both online data migration and replication in P2P environments.

A fast distributed slicing algorithm

Proceedings of the twenty-seventh ACM symposium on Principles of distributed computing - PODC '08, 2008

Slicing a distributed system involves partitioning the nodes into k equal-size subsets using a onedimensional attribute. A new gossip-based slicing algorithm that we call Sliver is proposed here; relative to alternatives, it converges more rapidly to an accurate solution and does a better job of tolerating churn. The low cost and simplicity of the solution makes it appropriate for use in a wide range of practical settings.

Dynamic Load Sharing in Peer-to-Peer Systems: When Some Peers Are More Equal than Others

IEEE Internet Computing, 2000

In the past few years, several DHT-based abstractions for peer-to-peer systems have been proposed. The main characteristic is to associate nodes (peers) with keys (objects) and to construct distributed routing structures to support an efficient location. These approaches address the load problem, and load balancing is achieved by moving the keys. However, the problem is still not properly covered. In this paper we present an analysis of structured peer-to-peer systems taking into consideration Zipf-like requests distribution. Based on our analysis, we propose a novel approach for load balancing relying on object popularity. Our approach is based on routing table reorganization in order to balance the lookup traffic load. We have implemented this approach in a Pastry-like system. The obtained results demonstrate a better balance of load, which can lead to improved scalability and performance.

On Optimal Replication Group Splits in P2P Data Stores Based on the Hypercube

2007

P2P data stores excel if availability of inserted data items must be guaranteed. Their inherent mechanisms to counter peer pop- ulation dynamics make them suitable for a wide range of application domains. This paper presents and analyzes the split maintenance opera- tion of our P2P data store. The operation aims at reorganizing replication groups in case operation of them becomes sub-optimal. To this end, we present a formal cost model that peers use to compute optimal points when to run performance optimizing maintenance. Finally, we present experimental results that validate our cost model by simulating various network conditions.

Load Sharing in Peer-to-Peer Networks using Dynamic Replication

2006

The peer-to-peer (P2P) architecture provides support for the next generation of information sharing applications. A difficult challenge faced by these systems in the presence of non-uniform data distribution and dynamic network conditions is load sharing. This paper addresses the problem of load sharing in P2P networks across heterogeneous super-peers. We propose two load sharing techniques that use data replication to improve access performance. In the first technique, called Periodic Push-based Replication (PPR), super-peers periodically send replicas of the most frequently accessed files to remote super-peers. This effectively reduces the hop count to fetch these files. The second technique, called On-Demand Replication (ODR), performs replication based on access frequency. By performing replication on-demand, ODR provides adaptability to changes in access behavior. Extensive testing have been conducted to study the performance of the proposed techniques. The results obtained demonstrate significant performance improvements through replication.

Dante: A self-adapting peer-to-peer system

2008

In this paper we introduce DANTE, an unstructured P2P system in which the topology of the underlying overlay network can be dynamically adapted to the system conditions. Such an adaption is performed by the peers in an autonomous manner. DANTE uses a simple search mechanism based on random walks that, combined with the topology adaptation, allows it to work in a very efficient way. We have evaluated how DANTE behaves in practice, showing that it successfully adapts to varying system load conditions.

Policies for Efficient Data Replication in P2P Systems

2013 International Conference on Parallel and Distributed Systems, 2013

This paper addresses the problem of maintaining replicated data in large scale P2P systems. Although this topic has been extensively studied in the literature, to maintain replicated data in this setting, in an efficient manner, still remains a significant challenge. This paper proposes novel policies to address this problem and evaluates their performance against different criteria, such as monitoring costs, data transfer costs, and load unbalance costs. We show that one of these new policies significantly outperforms previous work. Interestingly, this policy is based on a somehow counter-intuitive approach, that uses less reliable nodes to store the most accessed data items. The insights to derive this policy were obtained from an in depth analysis of existing solutions, that is also captured in the paper.