Accelerating network function virtualization

Network function virtualization (NFV) [50] is increasingly used to implement network operations traditionally implemented in customized ASICs. NFV employs commodity, general-purpose computer hardware located in a datacenter. General-purpose computing hardware has performance limitations limiting the scope of NFV. An emerging solution is to leverage programmable accelerators such as GPUs and FPGAs for NFV. However, traditional computer architecture research of NFV applications is challenging, due to the lack of NFV benchmark suites. This dissertation presents a set of candidates for NFV benchmark suite, based on analysis of most common NFV application. We then study NFV acceleration on GPUs. We identify overheads that are especially pronounced in Service Function Chains (SFC) that are common in NFV. This dissertation proposes GPUChain, a mechanism to enhance SFC acceleration on GPUs by moving the chaining capability onto the GPU. GPUChain avoids the significant latency overheads incurred by a CPU-centric SFC execution model. GPUChain achieves an average latency reduction of 44% and improves throughput by 168% versus a GPU supporting pre-registered kernels that are triggered to launch by an external device [110] while incurring only 0.007% area overhead.Applied Science, Faculty ofElectrical and Computer Engineering, Department ofGraduat