high performance computing on graphics processing units: hgpu.org

Advancements in hardware changed the bottleneck of modern database systems from disk IO to main memory access and processing power. Since the performance of modern processors is primarily limited by a fixed energy budget, hardware vendors are forced to specialize processors. Consequently, processors become increasingly heterogeneous, which already became commodity in the form of accelerated processing units or dedicated co-processors such as graphics processing units. However, building a robust and efficient query engine for such heterogeneous co-processor environments is still a significant challenge. Although the database community developed fast parallel algorithms for a large number of heterogeneous processors, we still require methods to use these processors efficiently during query processing. This thesis shows how we can build database management systems that efficiently use heterogeneous processors to reliably accelerate database query processing. Thus, we explore the design space of such co-processor-accelerated DBMSs to derive a generic architecture of such systems. Our investigations reveal that one of the most crucial problems in database engines running on heterogeneous processors is to decide which operator of a query should be executed on which processor. We refer to this as the operator placement problem. Common analytical modeling would greatly increase the complexity of a DBMS, because this complexity directly relates to the degree of heterogeneity of processors. Thus, we present a framework for hardware-oblivious operator placement called HyPE, which hides the processor heterogeneity from the DBMS by using statistical learning methods and efficiently balances the load between processors. Furthermore, we examine performance and scalability of query processing for co-processor-accelerated DBMSs in case common assumptions of co-processing techniques are not met. Our investigations show that co-processors can significantly slow down a DBMS when not used appropriately and develop approaches that avoid using co-processors when we expect a performance degradation. Throughout this thesis, we show the efficiency of our approaches by integrating them into our open source database engine CoGaDB.