UC Irvine

Data are among the most valuable assets in the modern world, and they have caused a revolutionary stage in human life. Nowadays, companies make knowledge-based decisions by analyzing a huge volume of data, super-scale data centers are used to process customers’ data to suggest products to them, government services rely on the data people provide to them, and there are many similar cases wherein data are used as an important asset. Data are originally stored in storage systems. To process data, application servers need to fetch the data from storage units, which imposes the cost of moving the data to the system. This cost has a direct relationship to the distance of the processing engines from the data, and this is the key motivation for the emergence of distributed processing platforms such as Hadoop, which bring the process closer to the data.

In-storage processing (ISP) pushes the “bring the process to data” paradigm to its ultimate boundaries by utilizing processing engines inside the storage units to process data. The architecture of modern solid-state drives (SSDs) provides a suitable environment for implementing such technology. Thus, this dissertation focuses on SSD architectures that are able to run user applications in-place, which are called computational storage devices (CSDs). In this dissertation, we propose CSD architectures and investigate the benefits of deploying CSDs for running different applications. This research uses a practical approach that includes building fully functional prototypes of the proposed CSD architectures, developing storage systems equipped with the CSDs, and running different benchmarks to investigate the benefits of deploying the CSDs in the systems. This research proposes two different CSD architectures, namely CompStor and Catalina.

These are the first CSDs to be equipped with a dedicated ISP engine for running user applications in-place that includes a quad-core ARM Cortex-A53 processor together with FPGA- and application-specific integrated circuit (ASIC) based accelerators. The proposed architectures run a full-fledged operating system inside, which provides a flexible environment for running a wide range of user applications in-place. The system-on-chip (SOC) based architecture of Catalina CSD, together with a software stack developed for seamless deployment of the CSD, makes it a platform for the implementation of different ISP concepts and ideas.

To the best of our knowledge, Catalina is the only ISP platform that can be seamlessly deployed in clusters to run distributed applications such as Hadoop MapReduce and message passing interface (MPI) based applications in-place without any modifications to the underlying distributed processing framework. We performed extensive experimental tests using several datasets on both CompStor and Catalina CSDs. The experimental results show up to 2.2x and 4.3x improvements in performance and energy consumption, respectively, for running Hadoop MapReduce benchmarks using Catalina CSDs and up to 5.4x and 8.9x improvements for running 1-, 2-, and 3-dimensional DFT algorithms due to the Neon SIMD engines inside Catalina. Additionally, using FPGA-based accelerators, Catalina CSDs can improve the performance and energy consumption of a highly demanding image similarity search application up to 11x and 7x, respectively.