What is it about?
We propose TOSCA, a novel dataflow architecture that improves performance through semi-centralized dynamic scheduling. Unlike conventional centralized schedulers that suffer from high control overhead, or decentralized approaches that lack global workload awareness, TOSCA combines global visibility with local agility through hardware-software co-design. TOSCA introduces a semi-centralized scheduling execution model with on-demand triggering, global workload sensing, global-aware planning, and low-latency task remapping. It further designs a dedicated task flow plane with semi-proactive processing elements (PEs) and a global-monitoring controller, enabling efficient coordination between local execution and global scheduling decisions. Through adaptive load balancing mechanisms that consider both scheduling cost and workload imbalance, TOSCA effectively improves parallelism and resource utilization. Evaluated on a wide range of representative applications, TOSCA achieves up to 1.52× performance improvement and 1.40× energy-efficiency improvement over state-of-the-art dataflow architectures.
Featured Image
Photo by Steve A Johnson on Unsplash
Why is it important?
Dataflow architectures provide high flexibility and massive parallelism, making them promising for emerging computing workloads. However, efficiently utilizing large-scale dataflow architectures requires effective dynamic scheduling to adapt to irregular workloads and runtime variations. Existing solutions face a fundamental trade-off: centralized scheduling provides global workload awareness but introduces significant communication and control overhead, while decentralized scheduling reduces overhead but may make suboptimal decisions due to limited visibility. TOSCA addresses this challenge by establishing a new semi-centralized scheduling paradigm that combines the strengths of both approaches. By coordinating global monitoring with local autonomous execution, TOSCA provides a scalable and efficient solution for improving workload balance and hardware utilization in future dataflow systems.
Perspectives
For future edge computing, the real challenge of dataflow architectures lies in how effectively they can balance global scheduling awareness and local execution autonomy to achieve scalable dynamic resource management.
Shantian Qin
Institute of Computing Technology, Chinese Academy of Sciences
Read the Original
This page is a summary of: TOSCA: Towards Semi-Centralized Scheduling in Dataflow Architectures via Hardware-Software Co-Design, ACM Transactions on Architecture and Code Optimization, July 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3833878.
You can read the full text:
Contributors
The following have contributed to this page
