Reconfigurable-Hardware Accelerated Stream Aggregation

High throughput and low latency stream aggregation is essential for many applications that analyze massive volumes of data in real-time. Incoming data need to be stored in a single sliding-window before processing, in cases where incremental aggregations are wasteful or not possible at all. However, storing all incoming values in a single-window puts tremendous pressure on the memory bandwidth and capacity. GPU and CPU memory management is inefficient for this task as it introduces unnecessary data movement that wastes bandwidth. FPGAs can make more efficient use of their memory but existing approaches employ only on-chip memory and therefore, can only support small problem sizes (i.e. small sliding windows and number of keys) due to the limited capacity. This thesis addresses the above limitations of stream processing systems by proposing techniques for accelerating single sliding-window stream aggregation using FPGAs to achieve line-rate processing throughput and ultra low latency. It does so first by building accelerators using FPGAs and second, by alleviating the memory pressure posed by single-window stream aggregation. The initial part of this thesis presents the accelerators for both windowing policies, namely, tuple- and time-\ua0based, using Maxeler\u27s DataFlow Engines\ua0(DFEs) which have a direct feed of incoming data from the network as well as direct access to off-chip DRAM. Compared to state-of-the-art stream processing software system, the DFEs offer 1-2 orders of magnitude higher processing throughput and 4 orders of magnitude lower latency. The later part of this thesis focuses on alleviating the memory pressure due to the various steps in single-window stream aggregation. Updating the window with new incoming values and reading it to feed the aggregation functions are the two primary steps in stream aggregation. The high on-chip SRAM bandwidth enables line-rate processing, but only for small problem sizes due to the limited capacity. The larger off-chip DRAM size supports larger problems, but falls short on performance due to lower bandwidth. In order to bridge this gap, this thesis introduces a specialized memory hierarchy for stream aggregation. It employs Multi-Level Queues (MLQs) spanning across multiple memory levels with different characteristics to offer both high bandwidth and capacity. In doing so, larger stream aggregation problems can be supported at line-rate performance, outperforming existing competing solutions. Compared to designs with only on-chip memory, our approach supports 4 orders of magnitude larger problems. Compared to designs that use only DRAM, our design achieves up to 8x higher throughput. Finally, this thesis aims to alleviate the memory pressure due to the window-aggregation step. Although window-updates can be supported efficiently using MLQs, frequent window-aggregations remain a performance bottleneck. This thesis addresses this problem by introducing StreamZip, a dataflow stream aggregation engine that is able to compress the sliding-windows. StreamZip deals with a number of data and control dependency challenges to integrate a compressor in the stream aggregation pipeline and alleviate the memory pressure posed by frequent aggregations. In doing so, StreamZip offers higher throughput as well as larger effective window capacity to support larger problems. StreamZip supports diverse compression algorithms offering both lossless and lossy compression to fixed- as well as floating- point numbers. Compared to designs using MLQs, StreamZip lossless and lossy designs achieve up to 7.5x and 22x higher throughput, while improving the effective memory capacity by up to 5x and 23x, respectively