Doctoral Thesis: The Locality-First Strategy for Developing Efficient Multicore Algorithms

Monday, December 13
10am

Star (32-D463) and also on zoom (please contact me for the link at hjxu@mit.edu)

Helen Xu

Abstract:
In order to scale applications on multicores up to bigger problems, software systems must be optimized simultaneously for parallelism and cache-friendliness in order to take full advantage of the multiple cores and the memory hierarchy, respectively. Optimizing for either of these features is notoriously difficult, however, and combining them only adds to the complexity.

This talk will contend that in order to create parallel algorithms for multicores that are theoretically and practically efficient, practitioners should use a locality-first strategy. That is, they should first understand and exploit locality as much as possible before introducing parallelism. As an example, the talk will cover dynamic graph processing as a case study for the locality-first strategy.

Details

  • Date: Monday, December 13
  • Time: 10am
  • Category:
  • Location: Star (32-D463) and also on zoom (please contact me for the link at hjxu@mit.edu)
Additional Location Details:

Thesis Supervisor: Prof. Charles E. Leiserson