Elastic Data Transfer Optimization with Hybrid Reinforcement Learning

Modern scientific data acquisition generates petabytes of data that must be transferred to geographically distant computing clusters. Conventional tools either rely on preconfigured sessions, which are difficult to tune for users without domain expertise, or they adaptively optimize only concurrency while ignoring other important parameters. We present \name, an adaptive data transfer method that jointly considers multiple parameters. Our solution incorporates heuristic-based parallelism, infinite pipelining, and a deep reinforcement learning based concurrency optimizer. To make agent training practical, we introduce a lightweight network simulator that reduces training time to less than four minutes and provides a $2750\times$ speedup compared to online training. Experimental evaluation shows that \name consistently outperforms existing methods across diverse datasets, achieving up to 9.5x higher throughput compared to state-of-the-art solutions.

💡 Research Summary

The paper “Elastic Data Transfer Optimization with Hybrid Reinforcement Learning” addresses the critical challenge of efficiently transferring petabyte-scale scientific data over high-performance networks (HPNs). It identifies key limitations in existing tools: static, pre-configured sessions that cannot adapt to dynamic network conditions, and modern adaptive solutions that optimize only for concurrency while neglecting other crucial parameters like parallelism and pipelining. This narrow focus leads to suboptimal performance for datasets with heterogeneous file size distributions.

To overcome these limitations, the authors propose EDT, an adaptive data transfer framework that employs a novel hybrid strategy to jointly optimize three fundamental parameters: parallelism, pipelining, and concurrency. Each parameter is handled with a tailored approach:

1. Parallelism is managed heuristically. Instead of a fixed number of streams per file, EDT uses a maximum chunk size. Each file is divided into ceil(file_size / chunk_size) parts, dynamically allocating more streams to larger files and avoiding unnecessary overhead for small files.
2. Pipelining is implemented as an “infinite” strategy. Data channel connections are kept open until a transfer is complete or paused by the controller. This prevents the TCP congestion window from resetting between small file transfers, eliminating control-channel idle time and significantly improving performance for datasets with numerous small files.
3. Concurrency, deemed the most impactful parameter, is optimized using a Deep Reinforcement Learning (DRL) agent based on the Proximal Policy Optimization (PPO) algorithm. The agent observes real-time network state (e.g., bandwidth, RTT) and learns a policy to determine the optimal number of concurrent file transfers, balancing utilization against system overhead and congestion.

A major innovation enabling the practical use of DRL is the introduction of a lightweight network simulator. Training a DRL agent directly on a production network is impractical, often requiring days. EDT’s simulator accurately emulates data transfer dynamics under various network conditions and file distributions, allowing the agent to be trained offline in less than four minutes—a 2750x speedup compared to online training. This makes the sophisticated ML-based optimization deployable in real-world scenarios.

The experimental evaluation demonstrates EDT’s superiority across diverse datasets (uniform medium files, many small files, few large files). It is compared against static configuration baselines and state-of-the-art adaptive concurrency-only optimizers. Results show that EDT consistently achieves the highest and most stable throughput. Most notably, it delivers up to 9.5x higher throughput than the best existing solutions by effectively leveraging the combination of all three optimization techniques. The hybrid design ensures robust performance regardless of file size distribution, while the efficient simulator makes advanced ML optimization feasible for production use, marking a significant step forward in intelligent data movement for high-performance computing.