Processing-in-Memory Technology for Knowledge Discovery Algorithms
Abstract
The goal of this project was to gain insight into whether processing-in-memory (PIM) technology can be used to accelerate the performance of link discovery (LD) algorithms, which represent an important class of emerging knowledge discovery techniques being used by DoD to identify complex, multi-relational patterns. To this end, we evaluate the mapping of LD algorithms to a PIM workstation-class architecture, the DIVA/Godiva hardware testbeds developed by USC/ISI. These hardware testbeds incorporate PIMs into the memory of a conventional processor. Our performance measurements on bandwidth benchmarks, StreamAdd and RandomAccess, show that our prototype PIMs offer increased memory bandwidth to the applications over the Itanium2, with a commensurate increase in performance. The raw performance measurements for two LD kernels show a slowdown on a single PIM, but our analysis shows a performance gain when the differences in clock speed and data scaling are taken into account.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 01, 2006
- Accession Number
- ADA456908
Entities
People
- Hans Chalupsky
- Jacqueline Chame
- Jafar Adibi
- Mary Hall
- Tim Barrett
Organizations
- University of Southern California