A QoS aware multicore hash scheduler for network applications

As the line speed of the network evolves at an unprecedented rate, a wide spectrum of network applications call for increasing processing density on network devices. The prevalence of multicore chips ameliorates the stress on processing power, but the QoS guarantee is often ignored. In addition, results of legacy QoS studies are difficult to apply to multicore web servers. Therefore, a multicore scheduler that incorporates QoS concerns is missing. As the network development moves towards cloud computing, we see an increasing importance of QoS guarantees on high performance multicore network appliances. In this paper, we propose a proportional share hash based scheduler, PS-HRW, which extends existing optimizations in multicore scheduling with QoS concerns. We address the network QoS requirement by assigning weights to each connection following the classic General Processor Sharing (GPS) theory. Based on our previous multicore scheduling studies, PS-HRW allocates computing resources based on the QoS requirement, such that the workload is balanced at the packet level, and the connection locality is maintained. To provide accurate QoS guarantee, PS-HRW allocates an integral number of cores first and then allocates the residuals using a partitioning theory. However, different from traditional simulation based approach, we target at two popular applications on modern network appliances: Deep Packet Inspection (DPI) and multimedia transcoding. In addition, we generalize the topology of different multicore architectures into a communication matrix and optimize PS-HRW to incorporate cache awareness. Essentially, PS-HRW schedules incoming traffic efficiently by balancing between connection locality, load balancing, core/cache topology and QoS guarantees.