high performance computing on graphics processing units: hgpu.org

As the base of the software stack, system-level software is expected to provide efficient and scalable storage, communication, security and resource management functionalities. However, there are many computationally expensive functionalities at the system level, such as encryption, packet inspection, and error correction. All of these require substantial computing power. What’s more, today’s application workloads have entered gigabyte and terabyte scales, which demand even more computing power. To solve the rapidly increased computing power demand at the system level, this thesis proposes using parallel graphics processing units (GPUs) in system software. GPUs excel at parallel computing, and also have a much faster development trend in parallel performance than central processing units (CPUs). However, system-level software has been originally designed to be latency-oriented. GPUs are designed for long-running computation and large-scale data processing, which are throughput-oriented. Such mismatch makes it difficult to fit the system-level software with the GPUs. This thesis presents generic principles of system-level GPU computing developed during the process of creating our two general frameworks for integrating GPU computing in storage and network packet processing. The principles are generic design techniques and abstractions to deal with common system-level GPU computing challenges. Those principles have been evaluated in concrete cases including storage and network packet processing applications that have been augmented with GPU computing. The significant performance improvement found in the evaluation shows the effectiveness and efficiency of the proposed techniques and abstractions. This thesis also presents a literature survey of the relatively young system-level GPU computing area, to introduce the state of the art in both applications and techniques, and also their future potentials.