FPGA-Based Systems for Stream Data Analytics and I/O Data Transformations

Abstract

The distributed nature of the cloud, regarding resource placement and application execution, incurs large data transfer. As data movement is unavoidable, it lies within the systems’ capacities to ensure its effectiveness. This thesis takes a step towards enhancing data movement with compute capabilities, more precisely with data analytics and privacy computational features, with the ultimate goal of making data movement more efficient and secure using Field Programmable Gate Arrays (FPGAs) based compute engines. From the data analytics perspective, a correlation coefficient compute engine is designed and analyzed. This selection is based on the widespread adoption of correlation across database and ML systems that require knowledge of the correlation relationship between the input data before processing it. The experimental evaluation shows that correlation computation offloaded to an FPGA-based smartNIC can sustain streams arriving at 100 Gbps over an RDMA network, while performing a single pass over the data and requiring an order of magnitude less time for computing compared to CPU or GPU designs. Maintaining the continuity in the data characterization approach, a streaming analytics compute engine is analyzed. The system objective is to sustain a 100 Gbps TCP/IP data rate while characterizing input data (cardinality, second frequency moment, frequency distribution) in a manageable space and with a single pass over the data. In addition to accomplishing its intended objective, the system deployment on a single FPGA can match the performance of 70 CPU cores. From the data privacy perspective, the performance benefits of offloading the I/O path data transformation operations (encryption and decryption) are analyzed, taking into account the requirements of a relational analytics engine (SAP HANA), while providing three levels of security. The findings indicate that substantial benefits can be achieved, both in terms of performance and freeing CPU resources, in the cloud environment, and protecting the data while it is at rest or transmitted over the network.