Infoscience

Thesis

Accelerators for Data Processing

The explosive growth in digital data and its growing role in real-time analytics motivate the design of high-performance database management systems (DBMSs). Meanwhile, slowdown in supply voltage scaling has stymied improvements in core performance and ushered an era of power-limited chips. These developments motivate the design of software and hardware DBMS accelerators that (1) maximize utility by accelerating the dominant operations, and (2) provide flexibility in the choice of DBMS, data layout, and data types. In this thesis, we identify pointer-intensive data structure operations as a key performance and efficiency bottleneck in data analytics workloads. We observe that data analytics tasks include a large number of independent data structure lookups, each of which is characterized by dependent long-latency memory accesses due to pointer chasing. Unfortunately, exploiting such inter-lookup parallelism to overlap memory accesses from different lookups is not possible within the limited instruction window of modern out-of-order cores. Similarly, software prefetching techniques attempt to exploit inter-lookup parallelism by statically staging independent lookups, and hence break down in the face of irregularity across lookup stages. Based on these observations, we provide a dynamic software acceleration scheme for exploiting inter-lookup parallelism to hide the memory access latency despite the irregularities across lookups. Furthermore, we propose a programmable hardware accelerator to maximize the efficiency of the data structure lookups. As a result, through flexible hardware and software techniques we eliminate a key efficiency and performance bottleneck in data analytics operations.

Fulltext

Related material