Skip to main content

PERMAVOST 2025 keynote: dataflow under the von Neumann machine

Talks & Insights · · 2 min read

Xian-He Sun gave the keynote at the 5th Workshop on Performance Engineering, Modeling, Analysis, and Visualization Strategy (PERMAVOST 2025), held with ACM HPDC 2025 at the University of Notre Dame on July 20, 2025. His talk, "Dataflow under the von Neumann Machine: A New Paradigm for Computing Systems", spoke directly to the workshop's focus on performance modeling.

A paradigm for data-centric computing​

Sun started from the shift that big data and AI have caused: computing has become data-centric, so data management is now the first priority of a computing system. Traditional I/O and storage paradigms, he argued, no longer meet that demand, and a new paradigm is needed that joins high-performance computing with the velocity, volume and variety of data that different AI workloads need. He called that paradigm dataflow under the von Neumann machine, or DataflowV.

The models behind DataflowV​

The talk presented DataflowV in three steps. The Concurrent-AMAT (C-AMAT) model quantifies the benefit of concurrent data access and shows how data locality relates to concurrency; it gives a way to measure and close the performance gap at each layer of the memory hierarchy. The Layered Performance Matching (LPM) framework then optimizes a memory system layer by layer. On that base Sun defined DataflowV formally and showed results from applying it, including the Hermes multi-tiered I/O buffering system.

Open questions for performance modeling​

For a workshop on performance engineering, the last part of the talk mattered most. Sun said C-AMAT and LPM are powerful but leave many basic questions open. Have we used modern hardware advances to their full potential? Should advanced memory systems be built around concurrent-aware locality, a model of locality that also accounts for concurrency? He presented these as research problems for the modeling community rather than settled answers.

The workshop has published Sun's slides.

Learn more​