ACA 2020 keynote: data flow under the von Neumann architecture
Xian-He Sun gave an invited keynote at the 13th CCF National Computer Architecture Conference (CCF ACA 2020) on August 13, 2020. The conference, held every two years, ran online from August 10 to 16. His talk, "Data flow under the von Neumann Computer Architecture", asked how to design a data-centric architecture for the big data era and introduced an approach he named dataflowV.
A data-centric look at von Neumann
Big data applications have changed what computers spend their time on, and Sun opened with the question of how to design a computer architecture around data rather than computation. Both the von Neumann architecture and the classic dataflow architecture are compute-centric: dataflow machines arrange the movement of data so as to maximize parallel computation.
Looking at von Neumann from the data side, Sun argued that the model itself is neutral between computing and memory. Its compute-centric reputation comes from history, from the decades in which processors were the scarce resource, not from the architecture.
Modeling memory and data movement
The talk then examined von Neumann memory systems from a data-centric point of view and introduced metrics and methods to model and optimize them. Sun's group has published several of these, including the Concurrent Average Memory Access Time (C-AMAT) model in IEEE Computer in 2014 and the Layered Performance Matching (LPM) methodology in IEEE TPDS in 2019.
In the talk, memory system performance is modeled with four types of memory cycles. A hierarchical analysis of those four cycle types, together with the cost of computing, gives a formal trade-off between computing and data movement.
DataflowV
From that trade-off Sun proposed data flow under the von Neumann computer architecture, dataflowV for short. DataflowV decides where and how to compute by taking the cost of data movement into account. It is not a new computer architecture but a data-centric way of implementing the von Neumann architecture. A complete dataflowV system is a hard goal, but Sun noted that each of its point solutions can help current systems right away. He presented the concept, the modeling and the design, early implementation results, and the open research questions for dataflowV and for memory systems in general.
