2022-061 - GCNScheduler: Complex Task Scheduling with Graph Convolutional Networks

Description:
  • Distributed computing
  • Edge-cloud processing
  • Astronomical observation processing

Abstract

USC researchers present GCNScheduler, a scheduler employing graph convolutional networks (GCNs) that can rapidly and efficiently schedule tasks of complex applications for a given objective. The technique integrates an inter-task data dependency structure with network settings into an input graph to rapidly and efficiently develop scheduling schemes. GCNScheduler performs well in large-scale settings where past scheduling heuristics fail.

Benefits

  • Allows for complex task scheduling of large-scale applications
  • Operates orders of magnitude faster than the HEFT and TP-HEFT algorithms
  • Trains quickly and performs well

Market Application

Efficient task scheduling is essential in complex graph-based applications such as edge-cloud computing and astronomical observation processing. Successful task scheduling both improves the utilization of computing resources and reduces costs and execution time; however, the applicability of previously proposed scheduling heuristics is limited by their long computation times. Machine learning techniques offer an opportunity to improve task scheduling methods for large-scale task graphs.

Publications

Kiamari, Mehrdad, and Bhaskar Krishnamachari. "GCNScheduler: scheduling distributed computing applications using graph convolutional networks." Proceedings of the 1st International Workshop on Graph Neural Networking. 2022. https://doi.org/10.1145/3565473.3569185

Stage of Development

  • Tested against HEFT and TP-HEFT algorithms
  • Available for licensing

Patent Information:

  • Title: Scheduling Distributed Computing Based on Computational and Network Architecture
  • App Type: Nationalized PCT
  • Country: United States
  • Serial No.: 18/714,060
  • Patent No.:  
  • File Date: 5/28/2024
  • Issued Date:  
  • Expire Date:  
  • Patent Status: Patent Pending