Measurement- and model-based performance evaluation and speed-up of communications of multiprocessor PC systems in high-speed networks

点击翻译按钮获取中文摘要

The MEMPHIS project is concerned with methods for performance evaluation and improvement of the packet- and protocol-processing in current multicore PC systems. In the context of the MEMPHIS project the keyword "PC systems" stands for x86-compatible commodity hardware (32-bit and 64-bit systems), which means it includes diverse standard systems from a desktop machine to a high-end server. There is a broad range of application for packet- and protocol-processing on these systems, e.g. for software routers, middlebox applications (NAT, firewall) or virtual switches, which interconnect virtual machines (VMs) in virtualized data centers. Especially through the increasing degree of virtualization in data centers and also virtualization trends in other areas (e.g. Network Function Virtualization, NFV), packet-processing in PC systems is a highly topical subject.In order to be able to use PC systems for fast data transmission within the limits of their fundamental performance limitations, drivers, network stacks and applications need to be adapted better than before to the hardware architecture and its potential for parallelization. Which modifications are necessary for that purpose is being investigated in the first, currently running phase of the MEMPHIS project and forms the starting point for the second phase, which is the concluding third year of the MEMPHIS project. In a first step, models have already been developed, which describe the interaction between hardware and software inside existing systems and reflect the resource consumption of individual processing steps. In parallel, measurements have been performed on real systems, to permit parametrization and validation of the models. In the currently running first phase of the project we try to identify the performance-enhancing factors and bottlenecks of existing systems using our models and the performed measurements. In doing so, the models allow us to to predict the effects of different software-side modifications.Now, in the second phase of the project, we plan to transfer the insights gained in the first phase and make them usable in practice. This includes the realization, implementation and evaluation of performance-enhancing measures. We wish to make an important contribution for the development of future networking drivers, operating systems and applications in this second phase of the MEMPHIS project. Currently, we see a high relevance of our research, since there is great potential for software-based networking components to replace expensive and inflexible hardware systems in a number of areas.