Abstract
Distributed Transactional Memory (DTM) can play a fundamental role in the coordination of participants in edge clouds as a support for mobile distributed applications. DTM emerges as a concurrency mechanism aimed at simplifying distributed programming by allowing groups of operations to execute atomically, mirroring the well-known transaction model of relational databases.
Introduction
Distributed Transactional Memory (DTM) provides a powerful abstraction for developers to manage concurrent data access in distributed systems. By ensuring atomicity, DTM simplifies the complexity of building distributed applications. However, traditional DTM frameworks often rely on full data replication, which becomes a significant bottleneck in Edge Cloud environments. In these geographically dispersed settings, the high latency of wide-area networks makes the synchronization required for full replication prohibitively expensive.
To address this, researchers are exploring partial replication, where data is partitioned into groups. While partial replication reduces synchronization overhead, standard distribution methods like random assignment often ignore geographical locality, forcing transactions to span multiple distant data centers. This paper evaluates dynamic policies for mapping data to Partial Replication Groups (PRGs) to optimize performance in edge environments.
Methodology and Experimental Setup
The authors propose a Dynamic Resource Mapping Service (DRMS) that manages the placement of data objects across PRGs. They evaluated four primary policies:
- Static: Objects remain in their group of origin (baseline for natural locality).
- Move to the Latest: Data migrates to a new group upon every request from that group.
- Conservative Group Polling: Data moves only if usage in a target group is 1.5x higher than the global average.
- Aggressive Group Polling: Data moves to the group where it is most frequently accessed.
These were compared against location-insensitive baselines: Random and Round-Robin. Evaluation was conducted using a real-world Twitter dataset from August 2016, focused on four Portuguese cities (Lisbon, Porto, Braga, and Faro). Each city represented a PRG, and hashtags within tweets served as transactional data objects.
Key Results and Analysis
Performance and Locality
The study found that location-aware policies significantly outperform traditional distribution methods. As shown in the evaluation of transaction spans, the proposed policies (Static and Polling) consistently involved fewer groups per transaction compared to Random or Round-Robin assignments.
While Random distribution required transactions to contact over 2 groups on average, the proposed policies maintained a span of approximately 1.1 groups. This demonstrates that leveraging geographical locality allows the majority of transactions to be completed within a single local data center.
Stability and Migration Overhead
A critical trade-off in dynamic mapping is the cost of data migration. While the "Move to the Latest" policy ensures maximum locality for the immediate next transaction, it induces massive network churn.
As illustrated above, Move to the Latest triggered an average of 140 migrations per 10% of the workload. In contrast, Aggressive Group Polling provided a stable "sweet spot," achieving a 5% improvement in local transactions over the Static policy while maintaining very low migration overhead. Interestingly, the Static policy proved remarkably resilient, as social media data often exhibits strong "born-in" locality—users tend to use hashtags relevant to their immediate geographic area.
Discussion and Future Work
The research highlights that while dynamic polling is effective, certain challenges remain. Global events (like the "VMAs" hashtag) create spikes in popularity across all regions, making such data difficult to localize. Furthermore, the current model treats data objects independently. Future research should explore "correlation-aware" mapping, where related objects are moved together to further reduce transaction spans. Additionally, moving from reactive to proactive, predictive policies could further enhance performance in highly mobile edge scenarios.
Summary updated