On Fri, 21 Jun 2019 20:31:21 +0700 (WIB) [email protected] wrote:
> So I recently try to implement physical OvS-DPDK with Intel NIC (1Gbps). > Then, I compared the performance between native OvS and OvS-DPDK. I found out > that there is not much difference in the performance, I also tried changing > the packet size (range from 64B - 1500B) and the result: OvS-DPDK is > 1.05~1.3x higher than native OvS. Meanwhile, Intel performance report (this > is one of them: > https://download.01.org/packet-processing/ONPS2.1/Intel_ONP_Release_2.1_Performance_Test_Report_Rev1.0.pdf > ) stated that OvS-DPDK performance is 10~12x higher than native OvS (in the > link given, it is on section 7.2.1: PHY-OVS-PHY). > > The difference between "my OvS" and "Intel OvS" is Intel used 10Gbps NIC (and > even try to reach 40Gbps by using 4x10Gbps NIC), instead of me that just used > 1Gbps NIC. So here is my question: > > 1) Is DPDK performance-gain more likely to happen with higher speed NIC, such > as 10Gbps NIC rather than 1Gbps or 100 Mbps NIC? > 2) And to prove the question above, is there any performance report regarding > OvS-DPDK vs native-OvS with 1Gbps or 100 Mbps NIC? > > Thank you, > > Abraham A 1Gbit NIC is slow enough that CPU can keep up, so polling like DPDK doesn't really help. The standard Linux kernel GRO also helps and can keep up at those speeds. Summary: There really is no performance benefit expected with DPDK with fast cpu's at 1Gbit or less.
