Abstract
P2P architecture is a viable option for enterprise backup. In contrast to dedicated backup servers, making backups directly on organization's workstations should be cheaper, more efficient, and more reliable. We present the architecture of a p2p backup system that uses pairwise replication contracts between a data owner and a replicator. Unlike standard p2p storage systems using a DHT, contracts allow our system to optimize replicas' placement depending on specific strategies, taking advantage of machine and network heterogeneity. We built a scientific prototype and ran experiments on 150 workstations (13% average availability) and 50 PlanetLab nodes. Our main conclusion is that it is possible to build an efficient and reliable backup system even on highly unreliable machines.
Introduction
Enterprise backup traditionally relies on expensive dedicated hardware or centralized cloud providers, often costing upwards of $1,000/TB annually. However, most organizations possess significant untapped resources: the unused disk space on existing workstations. The primary obstacle to utilizing this "wasted" space is the inherent unreliability of non-dedicated machines, which are frequently turned off, rebooted, or crashed.
Standard peer-to-peer (P2P) storage systems typically employ Distributed Hash Tables (DHTs) for data placement. While effective for lookup, DHTs are often too rigid for backup scenarios where machine heterogeneity (varying stability, speed, or geographic location) can be leveraged for optimization. This research introduces a flexible P2P backup architecture designed to thrive in such environments.
Methods and Experimental Setup
The authors developed NebuloStore, a prototype backup system utilizing a "mixed architecture." While a standard DHT is used at the control layer for peer discovery, the storage layer abandons DHTs in favor of pairwise replication contracts. These are bilateral agreements where one peer agrees to store a chunk of another's data for a specified duration.
Utility-Based Placement
Data placement is governed by a utility function that scores potential replicators based on performance (minimizing backup/restore times), resiliency (maintaining redundancy), and geographic dispersion (ensuring at least one "far" replica to protect against local disasters).
The system employs a hill-climbing optimization strategy where peers periodically "gossip" to find better storage partners. To handle high churn, "synchro-peers" act as temporary caches for control messages intended for offline neighbors.
Evaluation Datasets
The system was evaluated using real-world data traces and two distinct network environments:
- University Computer Labs: 150 workstations with a median availability of just 13%, representing a "worst-case" scenario.
- PlanetLab: 50 stable nodes (91% availability) across Europe to test geographic optimization.
- Student Data Traces: Real data-size distributions from university students were used to simulate realistic backup loads across different storage quotas.
Results and Discussion
The experiments yielded several key insights into the feasibility of P2P backup on unreliable hardware.
The Impact of Availability
The "cost of unavailability" was the most significant factor. In the stable PlanetLab environment, backups averaged 1.1 hours, whereas the 13%-available lab environment required 3.1 hours. This delay is primarily due to the "waiting time" required for both the data owner and the replicator to be simultaneously online.
Contract-Based Optimization
The transition from DHTs to contracts proved successful. The system effectively distributed storage loads proportionally to machine availability—more stable machines naturally took on more replicas. In PlanetLab tests, the utility function successfully placed "far" replicas at an average distance of 8.1 hops while keeping "near" replicas close for speed.
Reliability and Scalability
Despite the extreme unreliability of the lab workstations, the system successfully maintained data redundancy. Furthermore, the total bandwidth scaled linearly with the number of machines, demonstrating that a P2P approach can effectively eliminate the bottlenecks associated with central backup servers. The use of synchro-peers was essential; without them, over half of the control messages would have been lost due to peer churn.
Conclusion
This research demonstrates that reliable enterprise backup can be built using existing, unreliable workstations. By moving away from rigid DHTs toward flexible pairwise contracts, the NebuloStore prototype achieved efficient data protection and optimized placement. Future work includes implementing data deduplication and expanding the system to public P2P networks, which would require robust incentive mechanisms and security protocols.
Summary updated