HPCC 2025 keynote: from parallel computing to concurrent data access
Xian-He Sun gave a keynote at the 27th IEEE International Conference on High Performance Computing and Communications (HPCC 2025) in Exeter, UK, which ran from August 13 to 15, 2025. In "From Parallel Computing to Concurrent Data Access: The Dataflow under von Neumann Machine Approach", he argued that decades of work on faster computation have made the data access problem worse.
Speedup is not the problem anymore
Deep learning succeeds because it can process vast amounts of data, but computing systems are struggling to keep up, and that has brought back the memory-wall problem. Sun's argument at HPCC was pointed. The tools the field relies on, parallel computing, multicore processors and dataflow architectures, were all designed to speed up computation. None was designed for the memory wall, and each puts more pressure on memory systems that are already overloaded. Data access has become the main performance limit of computing, and he called for rethinking computing from a data-centric point of view.
Concurrent data access
The approach he proposed is dataflow under the von Neumann machine. It rests on two GRC models. The C-AMAT model measures the benefit of serving many data accesses at once and shows how locality and concurrency relate. The Layered Performance Matching (LPM) framework uses that measure to tune each layer of the memory hierarchy. Together they shift the target from parallel computation to concurrent data access, which is the move named in the talk's title.
Sun then discussed GRC's work in I/O systems, centered on the Hermes multi-tiered I/O buffering system, which uses LPM to decide how data moves between storage tiers.
Shared infrastructure
The talk closed with ideas for computer system designs for AI and big data, and with the role of shared research infrastructure. Sun pointed to StoreHub, a proposed community facility for storage research that GRC is planning with The Ohio State University under an NSF grant, as one way to speed progress.
Learn more
- HPCC 2025 keynotes, with the abstract (opens in a new tab)
- Concurrent Average Memory Access Time (IEEE Computer, 2014)
- LPM: A Systematic Methodology for Concurrent Data Access Pattern Optimization (IEEE TPDS, 2019)
- SSDBM 2025 keynote: the data-centric imperative
- Hermes project page
- StoreHub project page
- Xian-He Sun's profile
