Is the performance of x550-t2 inferior to x540-t2 ?

[SYama19]

Could you tell us whether the events described below are within assumptions ?

From server1 where DPDK was installed to server2 installed in the same way, we sent 14.88 million 64byte-length packets with pktgen. At first, we installed x540-t2 in server1 and server2, and next did x550-t2 in server1 and server2.

In the measurement, when the packet-length is 64 bytes, x540-t2 received all 14.88 million packets, while x550-t2 received only 4.82 million packets.

When the packet-length is 128 bytes, x540-t2 received 8.44 million packets, while x550-t2 received only 4.62 million packets.

These numbers show the performance of x550-t2 inferior to x540-t2 although, given the specifications, the results should be the opposite.

I don't know why it will result like this. Could someone let us know why this could happen ?

メッセージ編集者: Shin Yama：

I think that the diagram will compensate for lack of my writing skills.

[idata]

Hi Budweiser,

Thank you for posting in Wired Communities. All Intel® DPDK questions and technical problems including those regarding the Ethernet* Controllers should be reported through the Intel® Premier Support site http://premier.intel.com/premier http://premier.intel.com/premier or access your IBL account and clickthe Intel® Premier Support link to enter issues under the Product Name "Data Plane Development Kit (DPDK)", and you will be routed to our support team for them to better assist you. Thank you.Regards,Sharon

[SYama19]

Hi Sharon,

Thank you for your kind reply.

I 'm really pleased to hear about how to get better support,

and would like to make effective use of your website.

Best regards,

Budweiser

[idata]

Hi Budweiser,

You are welcome and glad to be of help. Regards,Sharon

[Brian_J_Intel1]

I work at Intel in the Networking Division as a Solutions Architect

I did some quick testing using both the X540 and X550 dual-port adapters and found that when I run the tests on the local NUMA node I get ~14.8Mpps but when I run the same tests using the Remote NUMA node the results drop down to ~10-11Mpps. I suggest checking to see what NUMA node the X550 adapter is running on and if your test configuration has it running on the remote NUMA node, change it to run on the Local and see if that helps your performance.

Here is a summary of my server and test configurations:

Server and Test Configurations

Server 1 -- nd-r720-5

04:00.0 Ethernet controller: Intel Corporation Ethernet Controller 10G X550T (rev 01)04:00.1 Ethernet controller: Intel Corporation Ethernet Controller 10G X550T (rev 01)42:00.0 Ethernet controller: Intel Corporation Ethernet Controller 10-Gigabit X540-AT2 (rev 01)42:00.1 Ethernet controller: Intel Corporation Ethernet Controller 10-Gigabit X540-AT2 (rev 01) EAL: PCI device 0000:04:00.0 on NUMA socket 0EAL: PCI device 0000:42:00.0 on NUMA socket 1 6: enp4s0f0: mtu 1500 qdisc noop state DOWN group default qlen 1000 link/ether a0:36:9f:f5:a7:b0 brd ff:ff:ff:ff:ff:ff7: enp4s0f1: mtu 1500 qdisc noop state DOWN group default qlen 1000 link/ether a0:36:9f:f5:a7:b1 brd ff:ff:ff:ff:ff:ff8: enp66s0f0: mtu 1500 qdisc noop state DOWN group default qlen 1000 link/ether b4:96:91:01:9d:04 brd ff:ff:ff:ff:ff:ff9: enp66s0f1: mtu 1500 qdisc noop state DOWN group default qlen 1000 link/ether b4:96:91:01:9d:06 brd ff:ff:ff:ff:ff:ff

Server 2 -- nd-r720-92

05:00.0 Ethernet controller: Intel Corporation Ethernet Controller 10G X550T (rev 01)05:00.1 Ethernet controller: Intel Corporation Ethernet Controller 10G X550T (rev 01)42:00.0 Ethernet controller: Intel Corporation Ethernet Controller 10-Gigabit X540-AT2 (rev 01)42:00.1 Ethernet controller: Intel Corporation Ethernet Controller 10-Gigabit X540-AT2 (rev 01) EAL: PCI device 0000:05:00.0 on NUMA socket 0EAL: PCI device 0000:42:00.0 on NUMA socket 1 6: enp5s0f0: mtu 1500 qdisc noop state DOWN group default qlen 1000 link/ether a0:36:9f:f5:a7:4c brd ff:ff:ff:ff:ff:ff7: enp5s0f1: mtu 1500 qdisc noop state DOWN group default qlen 1000 link/ether a0:36:9f:f5:a7:4d brd ff:ff:ff:ff:ff:ff8: enp66s0f0: mtu 1500 qdisc noop state DOWN group default qlen 1000 link/ether b4:96:91:01:9e:5c brd ff:ff:ff:ff:ff:ff9: enp66s0f1: mtu 1500 qdisc noop state DOWN group default qlen 1000 link/ether b4:96:91:01:9e:5e brd ff:ff:ff:ff:ff:ff

Test Details

OS version: Ubuntu 16.04.3 LTS (GNU/Linux 4.4.0-98-generic x86_64)

DPDK version: 17.08.0

Pktgen version: 3.4.2

Test 1 64byte-packet with pktgen to testpmd (forward-mode=mac) both running on Local NUMA

Test 2 64byte-packet with pktgen to testpmd (forward-mode=mac) both running on Remote NUMA

Sender

./app/x86_64-native-linuxapp-gcc/pktgen -c 0xfffc -n 2 -w 0000:xx:00.0 --socket-mem 512,512 --proc-type auto -- -P -m [x:x].0 -f Sequence.Input

Receiver

./testpmd -c 0xxx -n 1 -w xx:00.0 -- --forward-mode=mac --eth-peer=0,xx:xx:xx:xx:xx:xx –I

Case 1 - Intel® Ethernet CNA X540-T2

Te...

[SYama19]

Thank you very much for your detailed answer.

In fact, we also conducted both local NUMA and remote NUMA tests,

and each result was 4.8 Mpps at x 550-t2 in the mesurement.

So we looked for other causes.

After seeing your answer,updating the version of DPDK from 16.11.3 to 17.08 and testing it, the packet drop at x550 - t2 disappeared.

Without your answer, we think we didn't notice that the DPDK version is the cause.

Once again, thank you for carrying out the performance test.