Adaptive Data Management in Evolving Heterogeneous Hardware/Software Systems
Currently, database systems face two big challenges: First, the application scenarios become more and more diverse ranging from purely relational to graph-shaped or stream-based data analysis. Second, the hardware landscape becomes more and more heterogeneous with standard multi-core Central Processing Units (CPUs) as well as specialized high-performance co-processors such as Graphics Processing Unit (GPUs) or Field Programmable Gate Arrays (FPGAs).
Recent research shows that operators designed for co-processors can outperform their CPU counterparts. However, most of the approaches focus on single-device processing to speedup single analyses not considering overall system performance. Consequently, they miss hidden performance potentials of parallel processing across all devices available in the system. Furthermore, current research results are hard to generalize and, thus, cannot be applied to other domains and devices.
In this project, we aim to provide integration concepts for diverse operators and heterogeneous hardware devices in adaptive database systems. We work on optimization strategies not only exploiting individual device-specific features but also the inherent cross-device parallelism in multi-device systems. Thereby we focus on operators from the relational and graph domain to derive concepts not limited to a certain application domain. To achieve the project goals, interfaces and abstraction concepts for operators and processing devices have to be defined. Furthermore, operator and device characteristics have to be made available to all system layers such that the software layer can account for device specific features and the hardware layer can adapt to the characteristics of the operators and data. The availability of device and operator characteristics is especially important for global query optimization to find a suitable execution strategy. Therefore, we also need to analyze the design space for query processing on heterogeneous hardware, in particular with regards to functional, data and cross-device parallelism. To handle the enormous complexity of the query optimization design space incurred by the parallelism, we follow a distributed optimization approach where optimization tasks are delegated to the lowest possible system layer. Lower layers also have a more precise view on device-specific features allowing to exploit them more efficiently. To avoid interferences of optimization decisions at different layers, a focus is also set on cross-layer optimizations strategies. These will incorporate learning- based techniques for evaluating optimization decisions at runtime to improve future optimization decisions. Moreover, we expect that learning-based strategies are best suited to integrate device-specific features not accounted for by the initial system design, such as it is often the case with the dynamic partial reconfiguration capabilities of FPGAs.