Fast and Efficient Update Handling for Graph H2TAP

Offloading graph analytics to GPU yields significant performance speedups. In heterogeneous hybrid transactional/analytical graph processing (graph H 2 TAP), where each graph workload type is executed on the most suitable processor, transactions are executed on a CPU-based main graph and analytics are executed on a GPU-optimized graph replica. The problem that arises, as a result, is that updates by transactions on the main graph have to be particularly handled with respect to the graph replica. In this paper, we present a fast and efficient approach to this update handling problem, based on a delta store optimized for graphs. The delta store is a differential graph store that captures the transactional updates, which are later propagated to the graph replica so that analytical queries are executed on the most recently committed version of the graph in accordance with freshness requirements. Our approach ensures consistency be-tween the main graph and the replica. Our evaluation shows the performance advantage of our approach over existing HTAP approaches.