[Paper Review] Data Load Balancing in Heterogeneous Dynamic Networks

Seyed Iman Mirrezaei, Javad Shahparian|arXiv (Cornell University)|Feb 15, 2016

Peer-to-Peer Network Technologies10 references3 citations

TL;DR

This paper proposes a distributed, topology-aware load balancing algorithm for structured P2P overlay networks that uses virtual servers and proximity-based load reassignment to reduce transfer costs and accelerate convergence. By leveraging physical network proximity and a parametric tradeoff between load quality and transfer cost, the approach reduces bandwidth usage by 43% in simulations, with minimal impact from in-degree imbalance.

ABSTRACT

Data load balancing is a challenging task in the P2P systems. Distributed hash table (DHT) abstraction, heterogeneous nodes, and non uniform distribution of objects are the reasons to cause load imbalance in structured P2P overlay networks. Previous works solved the load balancing problem by assuming the homogeneous capabilities of nodes, unawareness of the link latency during transferring load, and imposing logical structures to collect and reassign load. We propose a distributed load balancing algorithm with the topology awareness feature by using the concept of virtual servers. In our approach, each node collects neighborhood load information from physically close nodes and reassigns virtual servers to overlay nodes based upon the topology of underlying network. Consequently, our approach converges data load balancing quickly and it also reduces the load transfer cost between nodes. Moreover, our approach increases the quality of load balancing among close nodes of overlay and it also introduces a new tradeoff between the quality of load balancing and load transfer cost among all overlay nodes. Our simulations show that our approach reduces the load transfer cost and it saves network bandwidth respectively. Finally, we show that the in-degree imbalance of nodes, as a consequence of topology awareness, cannot lead to a remarkable problem in topology aware overlays.

Motivation & Objective

• Address load imbalance in heterogeneous, dynamic P2P networks caused by non-uniform object distribution and node heterogeneity.
• Overcome limitations of prior centralized or homogeneous-node assumptions in load balancing solutions.
• Minimize load transfer cost by prioritizing physically close nodes for data migration.
• Maintain scalability and avoid single points of failure by eliminating logical control structures.
• Introduce a tunable tradeoff between load balancing quality and network cost through a parametric algorithm.

Proposed method

• Each node collects load information from physically proximate neighbors using practical internet coordinates (PIC) to estimate network distance.
• The algorithm uses virtual servers to represent load capacity and enables dynamic reassignment of virtual servers between nearby nodes.
• A parametric mechanism balances load based on node capacity and proximity, allowing tradeoffs between load quality and transfer cost.
• Load reassignment is performed only between nodes with low network latency, reducing bandwidth consumption and improving convergence speed.
• The system avoids logical structures for load collection, preserving scalability and decentralization.
• A reactive mechanism handles in-degree imbalance by triggering routing table updates when high-in-degree nodes become overloaded.

Experimental results

Research questions

• RQ1 How can load balancing in structured P2P overlays be made more efficient by leveraging underlying network topology?
• RQ2 To what extent does topology awareness reduce load transfer cost and improve convergence speed?
• RQ3 Can a fully distributed, decentralized load balancing approach avoid the scalability and failure issues of centralized models?
• RQ4 What is the impact of topology awareness on routing table in-degree imbalance in overlays like RAQNet?
• RQ5 How can a tunable tradeoff between load balancing quality and transfer cost be effectively implemented in practice?

Key findings

• The proposed topology-aware load balancing reduces load transfer cost by 43% compared to non-topology-aware approaches in the GT-ITM topology model.
• 50% of load transfers occur over links with average latency of 100 ms when topology-aware, compared to 280 ms without awareness, significantly reducing network congestion.
• Over 80% of load transfers occur over links with average latency below 200 ms under topology awareness, indicating efficient use of low-latency paths.
• The algorithm converges quickly due to prioritization of physically close nodes, minimizing inter-node communication overhead.
• In-degree imbalance caused by proximity-based routing does not lead to a significant problem, as confirmed by simulations on 10,000-node RAQNet networks.
• A reactive mechanism effectively mitigates in-degree imbalance by allowing high-in-degree nodes to trigger routing table updates without increasing message path length.