[Paper Review] GCNScheduler: Scheduling Distributed Computing Applications using Graph Convolutional Networks

Mehrdad Kiamari, Bhaskar Krishnamachari|arXiv (Cornell University)|Oct 22, 2021

Advanced Graph Neural Networks37 references4 citations

TL;DR

GCNScheduler proposes a graph convolutional network (GCN)-based scheduler that integrates task graph dependencies and network resource characteristics into a unified input graph to enable fast, scalable task scheduling in distributed systems. It achieves makespan and throughput performance comparable to state-of-the-art heuristics like HEFT and TP-HEFT, but with orders-of-magnitude faster scheduling times—e.g., 4ms for 50-node graphs vs. 1500s for HEFT.

ABSTRACT

We consider the classical problem of scheduling task graphs corresponding to complex applications on distributed computing systems. A number of heuristics have been previously proposed to optimize task scheduling with respect to metrics such as makespan and throughput. However, they tend to be slow to run, particularly for larger problem instances, limiting their applicability in more dynamic systems. Motivated by the goal of solving these problems more rapidly, we propose, for the first time, a graph convolutional network-based scheduler (GCNScheduler). By carefully integrating an inter-task data dependency structure with network settings into an input graph and feeding it to an appropriate GCN, the GCNScheduler can efficiently schedule tasks of complex applications for a given objective. We evaluate our scheme with baselines through simulations. We show that not only can our scheme quickly and efficiently learn from existing scheduling schemes, but also it can easily be applied to large-scale settings where current scheduling schemes fail to handle. We show that it achieves better makespan than the classic HEFT algorithm, and almost the same throughput as throughput-oriented HEFT (TP-HEFT), while providing several orders of magnitude faster scheduling times in both cases. For example, for makespan minimization, GCNScheduler schedules 50-node task graphs in about 4 milliseconds while HEFT takes more than 1500 seconds; and for throughput maximization, GCNScheduler schedules 100-node task graphs in about 3.3 milliseconds, compared to about 6.9 seconds for TP-HEFT.

Motivation & Objective

• Address the scalability limitations of traditional heuristic schedulers like HEFT and TP-HEFT, which become infeasible for large task graphs due to high computational overhead.
• Enable fast, real-time scheduling in dynamic distributed environments such as IoT and edge-cloud systems where task graphs are complex and frequently re-scheduled.
• Leverage graph neural networks to model inter-task dependencies and network heterogeneity simultaneously, capturing structural and performance-aware relationships in task graphs.
• Achieve high scheduling efficiency without sacrificing optimization quality, particularly for makespan minimization and throughput maximization objectives.
• Demonstrate generalization across diverse workloads, including synthetic and real-world perception applications, to validate robustness and practicality.

Proposed method

• Construct a heterogeneous input graph combining task nodes (with execution cost features), communication edges (with bandwidth and latency features), and compute machine nodes (with processing speed features).
• Apply a spatial-based graph convolutional network (GCN) that aggregates neighborhood information through message-passing to generate task-specific node embeddings.
• Use a multi-layer GCN with ReLU activation and linear transformation to learn task-to-machine assignment policies from input graph representations.
• Train the GCN end-to-end using supervised imitation learning, where ground-truth schedules from HEFT and TP-HEFT serve as supervision signals.
• Optimize for two objectives: makespan minimization (via HEFT supervision) and throughput maximization (via TP-HEFT supervision), with separate training pipelines.
• Employ a softmax layer on final GCN embeddings to output a probability distribution over machine assignments for each task, enabling differentiable and efficient inference.

Experimental results

Research questions

• RQ1 Can a GCN-based scheduler achieve scheduling performance comparable to classic heuristics like HEFT and TP-HEFT while significantly reducing scheduling time?
• RQ2 How well does GCNScheduler generalize across different task graph sizes, from small to large-scale graphs (e.g., up to 5,000 tasks)?
• RQ3 To what extent can GCNScheduler maintain high throughput and low makespan across real-world applications such as face recognition and gesture recognition?
• RQ4 How does GCNScheduler’s inference speed scale with increasing task graph complexity compared to traditional schedulers?
• RQ5 Can the GCN model effectively learn from heuristic-based schedulers and generalize to unseen task graphs without retraining?

Key findings

• For makespan minimization, GCNScheduler schedules 50-node task graphs in approximately 4 milliseconds, compared to over 1500 seconds for HEFT, representing a 375x speedup.
• For throughput maximization, GCNScheduler schedules 100-node task graphs in 3.3 milliseconds, compared to 6.9 seconds for TP-HEFT, achieving a 2,100x speedup.
• In large-scale settings (e.g., 3,500–5,000 tasks), GCNScheduler completes scheduling in under 90 milliseconds, while TP-HEFT is infeasible due to excessive runtime.
• GCNScheduler achieves throughput performance slightly better than TP-HEFT on real perception applications (face, pose, and gesture recognition), with 2–3 orders of magnitude faster scheduling times.
• The model achieves a scheduling accuracy of approximately 95% on benchmark task graphs, indicating strong generalization and reliability.
• On real-world perception workloads, GCNScheduler reduces scheduling time from 87–290 milliseconds (TP-HEFT) to 0.488–0.560 milliseconds, demonstrating practical viability in real-time systems.