Researchers at Purdue University have presented GTX, a new system designed to efficiently manage large-scale dynamic graphs while supporting high-throughput read-write transactions and competitive graph analytics. This invention solves the issue which current transactional graph systems have with handling temporal localities and hotspots, two common features of real-world graphs.
Notably, managing such graphs is vital in many applications, such as fraud detection, recommendation systems, and graph neural network training. However, the current systems often use coarse-grained concurrency control mechanisms and often underperform when dealing with frequent update situations.
In response, the Purdue team’s proposed GTX system is a latch-free write-optimised transactional graph data system. It uses atomic operations to eliminate latches, resulting in the application of a delta-based multi-version storage system. It also implements a hybrid transaction commit protocol.
Uniquely, GTX incorporates a delta-chain index in its design, allowing it to manage concurrency control and support efficient edge lookups at the delta-chain level. The system was expressly designed to adapt to temporal localities and hotspots and maintain high throughput read-write transactions whilst still providing competitive graph analytics performance.
Taking a closer look at the architecture, GTX features a latch-free adjacency-list based graph store and a transaction manager that includes a concurrency control protocol. The system employs a multi-version delta store, where every delta captures vertex or edge operations, and this enables efficient access and updates.
The system manages concurrent transactions and analytics in harmony by controlling them at the delta-chain level and applying a hybrid group commit protocol. All these integrated systems increase overall throughput, providing an adaptive concurrency control based on workload history.
The researchers have tested GTX as a graph library prototype, using both real-world and synthetic datasets. The results showcase GTX’s impressive capability in handling real-world power-law graphs featuring temporal localities and hotspots, whilst sustaining millions of transactions per second and providing competitive graph analytics performance.
To summarize, the inventive GTX system presents a significant way forward in managing dynamic graphs with high arrival rates of updates, temporal localities, and hotspots. It outperforms existing systems regarding transaction throughput and robustness, across multiple workloads. The system’s ability to adapt to temporal localities and hotspots whilst maintaining vital graph analytics performance makes it a promising solution for many applications that rely on efficient graph management and analysis.
