Data Management Lab

Abstract:

Modern data-intensive systems perform complex analytical tasks on large datasets that keep growing at superlinear rates. Prevailing system architectures mandate that persistent data is trans- ferred across the whole memory hierarchy to the host to be processed there. Data movement limits the system performance and impacts scalability and resource consumption inversely. Yet, the emergence of intelligent storage/memory technologies and the ability to offload processing close to data creates new opportunities, as data movement is performed on-device much better performance and lower overall impact on processing. However, to date the decision of which operations to offload has been mostly hard-coded in near data processing DBMS. In this paper we propose hybridNDP in an attempt to automate offloading decisions given an ad hoc query. The core idea is to split queries into host- and on-device processing parts and enable cooperative intervention-free execution. To this end we propose a cost-model to determine potential splits and a cooper- ative execution model. We evaluate hybridNDP with nKV and the Join-Order Benchmark. Our findings indicate that through the offloading and execution scheme hybridNDP outperforms traditional host-only executions on various queries by up to 4.2x.