Bitwise Decomposed Data Processing
To address the problem of wasted resources caused by increasingly heterogeneous hardware, we proposed a generic processing paradigm for bitwise distributed relational data. Based on classic relational algebra, the paradigm is simple, yet powerful enough to a) provide a generic framework for efficient cross device processing of relational as well as array-data, b) allows architecture specific optimizations of individual operations and c) permit the parallel execution of operations on the available devices. This paradigm prescribes the multi-stage (Approximate & Refine) execution of relational operators. This leads to significant performance improvements (up to 8 times faster query evaluation than classic MonetDB) when handling large relational, as well as, array databases in a CPU/GPU co-processing setup. Additionally, it provides the possibility to compute a fast approximation of the query result before it is calculated at no additional costs. To back the paradigm, we introduced efficient algorithms for the arising problems and proved its feasibility by implementing it in MonetDB.
While BWD was a bonus result that was not necessary to support one of the complex TELEIOS cases, we demonstrated the performance benefits of a BWD-backed evaluation of SciQL queries in simpler benchmarks. We believe that the achieved performance can be generalized to future applications with stronger performance requirements. In addition, we plan to enhance the applicability of the paradigm to new applications like complex joins as well as other hardware setups like SSD/HDD combinations.