Reproduce “A Data-Driven Dynamic Execution Orchestration Architecture” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhenyu Bai, Pranav Dangi, Rohan Juneja, Zhaoying Li, Zhanglu Yan, Huiying Lan, Tulika Mitra
A data-driven dynamic execution orchestration architecture, Canon, integrates compile-time and runtime mechanisms to efficiently process both regular and irregular workloads. The architecture demonstrates competitive performance with specialized accelerators for their domain tasks and effectively manages general-purpose and irregular data, often outperforming CGRAs for parallel workloads and matching specialized sparse accelerators.
Key insights
- A hybrid orchestration approach, combining compile-time analysis for regular patterns and data-driven runtime FSMs for irregularities, can effectively bridge the specialization-flexibility gap in parallel architectures.
- Decoupling control from compute via row-level orchestrators and utilizing a time-lapsed SIMD instruction propagation scheme significantly reduces control overheads, enabling lightweight PEs and scalable performance.
- Strategically integrated dual-ported scratchpads at each Processing Element are critical for efficient data reuse, load balancing, and amortizing dynamic irregularities in sparse and irregular workloads.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.