Designing Near-data-processing solutions for Data-intensive Applications
Abstract:
Nowadays, data movement dominates the execution of applications with large memory footprints, due to the strict separation of computing devices (i.e., CPU) and data storage devices (i.e., main memory or disk). To overcome data movement overhead, the computer architecture community has been investigating various data-centric architectures, which are broadly referred to as Near-data-processing (NDP) technologies. NDP moves computation as close to data as possible, rather than fetching data from a remote location to processor cores. This Ph.D. dissertation proposal first reviews our previous effort on NDP accelerator designs that address the bottleneck stages in various data-intensive applications such as bioinformatics, exact pattern matching kernels, and database analytic queries, leveraging a wide selection of data storage technologies such DRAM and SSD. In addition, we notice many NDP designs focus on performance improvement over conventional architecture, while the potential security threats in such systems have been largely understudied. We describe a hardware supply chain attack against a NDP neural network accelerator that stealthily leaks model parameters. In our upcoming endeavor that further demonstrates the effectiveness of the NDP approach, we plan to leverage the efficient pattern-matching capability of our prior work, to develop an associative processor that executes arithmetic computation in a highly-parallel manner.
Committee:
- Felix Lin, Chair, CS/SEAS/UVA
- Kevin Skadron, Advisor, CS/SEAS/UVA
- Ashish Venkat, Co-advisor, CS/SEAS/UVA
- Sandhya Dwarkadas, CS/SEAS/UVA
- Mircea Stan, ECE/SEAS/UVA
- Adwait Jog, CS/SEAS/UVA