Efficient low-latency packet processing using On-GPU Thread-Data Remapping

摘要

Graphics processing units are widely-used for packet processing acceleration in both physical and virtual networks. However, real-life packets come in highly-divergent sizes, causing severe GPU control flow divergence. Previous solutions rely on CPU preprocessing to reduce divergence, but it forbids the more efficient NIC-GPU packet streaming as packet batches have to stop completely at host machine. To fully utilize both GPU and PCIe resources, we propose Blink as a GPU modular software router. Instead of CPU pre-processing, the Blink router uses On-GPU Thread-Data Remapping to reduce divergence, and our novel Cross-Iteration Thread Event Signaling mechanism filters unnecessary inter-thread synchronization, doubling the performance gain achieved by traditional solution. Serving as a TCP/IP router with Deep Packet Inspection (DPI) firewall, Blink can sustain processing throughput of 31.5 GBit/s over a PCIe bandwidth of 32 GBit/s. Given a certain bandwidth, Blink reduces processing latency at least by half compared with other works.