The Case For Heterogeneous HTAP

Raja Appuswamy, Manos Karpathiotakis, Danica Porobic, Anastasia Ailamaki
2017 Conference on Innovative Data Systems Research  
Modern database engines balance the demanding requirements of mixed, hybrid transactional and analytical processing (HTAP) workloads by relying on i) global shared memory, ii) system-wide cache coherence, and iii) massive parallelism. Thus, database engines are typically deployed on multi-socket multi-cores, which have been the only platform to support all three aspects. Two recent trends, however, indicate that these hardware assumptions will be invalidated in the near future. First, hardware
more » ... endors have started exploring alternate non-cache-coherent shared-memory multi-core designs due to escalating complexity in maintaining coherence across hundreds of cores. Second, as GPGPUs overcome programmability, performance, and interfacing limitations, they are being increasingly adopted by emerging servers to expose heterogeneous parallelism. It is thus necessary to revisit database engine design because current engines can neither deal with the lack of cache coherence nor exploit heterogeneous parallelism. In this paper, we make the case for Heterogeneous-HTAP (H 2 TAP), a new architecture explicitly targeted at emerging hardware. H 2 TAP engines store data in shared memory to maximize data freshness, pair workloads with ideal processor types to exploit heterogeneity, and use message passing with explicit processor cache management to circumvent the lack of cache coherence. Using Caldera, a prototype H 2 TAP engine, we show that the H 2 TAP architecture can be realized in practice and can offer performance competitive with specialized OLTP and OLAP engines.
dblp:conf/cidr/AppuswamyKPA17 fatcat:4dzw4rtkurazpgidsle3756mzy