Towards a Centralized Scheduling Framework for Communication Flows in Distributed Systems

The overall performance of a distributed system is highly dependent on the communication efficiency of the system. Although network resources (links, bandwidth) are becoming increasingly more available, the communication performance of data transfers involving large volumes of data does not necessarily improve at the same rate. This is due to the inefficient usage of the available network resources. A solution to this problem consists of data transfer scheduling techniques, which manage and allocate the network resources in an efficient manner. In this paper we present several online and offline data transfer optimization techniques, in the context of a centrally controlled distributed system.

💡 Research Summary

The paper addresses a fundamental bottleneck in modern distributed systems: despite the ever‑increasing availability of network links and bandwidth, large‑scale data transfers often fail to reap proportional performance gains because the underlying network resources are allocated inefficiently. To remedy this, the authors propose a centrally controlled scheduling framework that orchestrates communication flows across the entire system. The framework consists of a global controller that continuously gathers state information (available bandwidth, latency, packet loss, queue occupancy) from every node and link, and then decides how to route and allocate bandwidth for each incoming transfer request.

Two complementary optimization strategies are presented. The first is an online scheduler designed for real‑time arrivals. It assigns a composite weight to each request based on urgency, payload size, and the current congestion level of the candidate paths. A priority queue together with a dynamic re‑balancing routine ensures that high‑priority jobs are dispatched promptly while the overall load remains balanced. The second is an offline scheduler that assumes a known workload ahead of time. Here the authors formulate an integer linear programming (ILP) model whose objective is to minimize the sum of end‑to‑end transfer delays subject to constraints on link capacities, precedence relationships among jobs, and fairness requirements. Because solving the ILP exactly is computationally prohibitive for large clusters, they develop a suite of heuristics—Lagrangian relaxation, problem decomposition, and greedy rounding—that produce near‑optimal schedules within a time frame compatible with practical deployment.

Fairness across multiple tenants or applications is explicitly enforced. The framework monitors three fairness metrics: a guaranteed minimum bandwidth per tenant, a maximum tolerable latency, and a long‑term average utilization ratio. When any metric deviates from its target, the controller triggers a resource re‑allocation to prevent starvation or monopolization.

The authors validate their approach through both simulation and real‑world experiments. In a simulated environment with diverse topologies (fat‑tree, random mesh) and workloads (bulk data replication, iterative machine‑learning model updates), the centralized scheduler reduces average transfer latency by more than 30 % and improves overall network utilization by roughly 25 % compared with conventional distributed schedulers such as Hadoop YARN and Spark’s built‑in scheduler. A physical testbed consisting of 100 nodes interconnected by 10 Gbps links confirms these gains under realistic conditions, demonstrating that the framework scales to handle terabyte‑scale transfers without becoming a performance bottleneck itself.

Scalability of the central controller is a key concern. To avoid a single point of overload, the paper proposes a hierarchical control architecture: a top‑level global controller defines system‑wide policies, while a set of regional (edge) controllers manage local subsets of nodes, performing fine‑grained scheduling decisions based on the global policy. This decomposition allows the framework to be deployed in large cloud‑edge hybrids where thousands of nodes may be involved.

In conclusion, the work shows that a centrally coordinated scheduling mechanism—augmented with both online heuristics and offline ILP‑based optimization—can substantially improve communication efficiency in distributed systems. The authors suggest future directions such as integrating machine‑learning predictions of traffic patterns into the scheduler, extending the model to support multipath transport, and coupling the framework with network function virtualization to further enhance flexibility and performance.