High-speed stateful packet classifier based on TSS algorithm optimized for off-chip memories

Abstract

We present a modular out-of-order architecture for stateful packet classification. The architecture uses DDR4 SDRAM memory to store rules and their state information to support millions of rules. The memory access pattern generated by network traffic significantly degrades the performance of the DDR4. Our architecture contains a cache and aggregation queues to negate this effect. Additionally, the memory subsystem supports a read cancellation and uses an out-of-order pipeline to maximize the main memory’s effectiveness further. The rule set update is implemented as a non-blocking operation and can be interleaved with lookup operations without any performance decrease, leading to the same execution time for rule update and rule lookup. The architecture is optimized for the modern datacenter’s network traffic and a small on-chip memory footprint, making it suitable as an accelerator for the Open vSwitch. As a result, our novel architecture configured with 1 million exact match rules can process traffic up to 202 Gbit/s (300 Mp/s) in average case and 51 Gbit/s (76 Mp/s) in the worst case with the use of a common dual-channel 64 bit DDR4-2666 MHz. It uses fewer FPGA resources (excluding cache memory) than the well-known de facto industry standard Xilinx MIG DDR4 controllers. Our proposed architecture enables commodity FPGA cards commonly equipped with DDR4 to process 100 Gbit/s which results in a significant cost reduction of a 100G SmartNICs.