|Name:||TiDA: High-Level Programming Abstractions for Data Locality Management|
|Time:||Tuesday, June 21, 2016
12:00 pm - 12:30 pm
|Breaks:||12:30 pm - 01:45 pm Lunch|
|Speaker:||Didem Unat, Koç University|
|Abstract:||Contemporary HPC processor designs are moving towards massively parallel chips. This, combined with the relatively high energy costs for data movement compared to computation, gives paramount importance to data locality management in programs. Programming models play a crucial role in providing the necessary tools to express locality and minimize data movement, while also abstracting complexity from programmers. We propose three simple yet powerful programming abstractions, tiles, layout and loop traversal, that address the dual challenges of data locality and increased parallelism for emerging parallel computing systems. TiDA implements these abstractions by elevating tiling to the programming model to expose high degrees of parallelism through domain decomposition. We demonstrate how TiDA automates both cache-locality and topology optimizations with minimal coding effort on current and future NUMA node architectures.
Didem Unat, Koç University; Tan Nguyen, LBNL; Weiqun Zhang, LBNL; Muhammed Nufail Farooqi, Koç University; Burak Bastem, Koç University; George Michelogiannakis, LBNL; Ann Almgren, LBNL; John Shalf, LBNL