ACM TURC 2019 keynote: data access of big data
Xian-He Sun gave the third keynote of the SIGMOBILE symposium at the ACM Turing Celebration Conference, China (ACM TURC 2019), in Chengdu on May 19, 2019. His talk, "Data Access of Big Data: Solving the system bottleneck", made the case that storage access limits big data systems, even mobile ones, and presented Hermes as GRC's response.
The I/O gap
Disk-based storage has improved far more slowly than memory and networks. The result is a wide gap between how fast a system can compute and how fast it can read and write data. Storage subsystems have answered by adopting new technologies, such as burst buffers and non-volatile memory, and by adding layers to the memory and storage hierarchy. Sun called the result a deep memory and storage hierarchy (DMSH).
More layers give more room for performance, but they also make data movement harder. An application now has to decide which tier should hold which data, and when to move it. Most applications cannot make those decisions well, so much of the hierarchy's potential goes unused.
The first part of the talk was about measurement: storage data access is the performance bottleneck that matters most, and that holds even in distributed mobile environments, the focus of the SIGMOBILE audience.
Hermes
The second part presented Hermes, a large NSF-supported software project that GRC leads with The HDF Group. Hermes is an intelligent, multi-tiered, dynamic and distributed I/O buffering system. It places data across the tiers of a DMSH so that applications see faster I/O without managing the tiers themselves.
Sun described two directions for the work. The first extends HPC I/O stacks so that memory and parallel I/O systems act as one integrated system. The second extends HDF5, the data format and library used widely in science, so that it can learn what an application needs and optimize data placement for it. He covered the Hermes design and implementation, what sets it apart from earlier caching layers, the open challenges, and early results. The design was first published at ACM HPDC 2018 by Anthony Kougkas, Hariharan Devarajan and Sun.
