Benchmarks

System under test

CPU: 2x Intel(R) Xeon(R) Gold 6348 CPU @ 2.60GHz
NIC: 2x Intel(R) E810-2cqda2 (100G 2 ports)
PCIe: 4th gen
Total network bandwidth: 400 Gbps
Distro: Ubuntu 20.04 LTS
Kernel: 5.17
Driver: ice-1.9.11

Traffic generators

CPU: 2x Intel(R) Xeon(R) Gold 6242R CPU @ 3.10GHz
NIC: 2x Intel(R) E810 (100G 1 of 2 ports connected)
PCIe: 3rd gen
Total network bandwidth: 200 Gbps
Distro: Ubuntu 20.04 LTS
Kernel: 5.17
Generator: Cisco T-Rex

We use 2 servers for the generation.

Traffic patterns

• Line rate - maximum practical throughput
• Base - base performance of XDP action, no regular expressions attached
• Literal - lookup for a constant sub-string inside of payload. Regex: /private/s; Corpus: printable characters
• Full payload scan - search with access to the whole payload. Regex /pri.*ate/sH; Corpus: printable characters
• Real-life - Our real-life regular expressions. 10 regexes in parallel, backtracking, search from the payload beginning.

We have evaluated XDP_DROP and XDP_TX actions as primerally used in our network.

Results

XDP_DROP

Line rate

pkt size	mpps	gbps	cpu usage
1500	32.0	384.0	n/a
1000	48.0	379.0	n/a
512	88.0	368.0	n/a
256	176.0	356.0	n/a
128	328.0	346.0	n/a
64	568.0	304.0	n/a

Base

pkt size	mpps	gbps	cpu usage
1500	32.0	384.0	6%
1000	48.0	379.0	8%
512	88.0	349.5	16%
256	162.0	340.0	28%
128	307.0	314.0	51%
64	430.0	221.0	75%

Literal

pkt size	mpps	gbps	cpu usage
1500	32.0	384.0	14%
1000	48.0	379.0	20%
512	88.0	349.5	34%
256	165.0	338.0	68%
128	244.0	250.0	98%
64	257.0	128.0	98%

Full payload scan

pkt size	mpps	gbps	cpu usage
1500	32.0	384.0	33%
1000	48.0	379.0	48%
512	88.0	349.5	80%
384	100.0	305.0	96%
256	108.0	220.0	98%
128	108.0	110.0	98%
64	211.0	108.0	98%

Real-life

pkt size	mpps	gbps	cpu usage
1500	32.0	384.0	15%
1000	48.0	379.0	22%
512	88.0	349.5	38%
256	161.0	330.0	74%
128	208.0	213.0	98%
64	220.0	112.0	99%