DPDK & QPI performance issue in Romley platform.

Hi~

I have a question about DPDK & QPI performance issue in Romley platform. Recently, I use DPDK example, l2fwd, to test DPDK's performance in my Romley platform. When I try to do the test, crossing used CPU, I find the performance dramatically decrease. Is it true? Or any method can prove the phenomenon?

In my opinion, there should be no this kind of issue here due to QPI have enough bandwidth to deal the kinds of case. Thus, I am so amaze in our results and can not explain it.
Could someone can help me to solve this problem.

Thank a lot!

My testing environment describe as below:

Platform:         Romley
CPU:                E5-2643 * 2
RAM:               Transcend 8GB PC3-1600 DDR3 * 8
OS:                 Fedora core 14
DPDK:            v1.3.1r2, example/l2fwd

Slot setting:
SlotA is controled by CPU1 directly. SlotB is controled by CPU0 directly.

DPDK pre-setting:
a. BIOS setting:
HT=disable
b. Kernel paramaters
isolcpus=2,3,6,7
default_hugepagesz=1024M
hugepagesz=1024M
hugepages=16
c. OS setting:
service avahi-daemon stop
service NetworkManager stop
service iptables stop
service acpid stop
selinux disable

Example program Command:
a. SlotB(CPU0) -> CPU1
#>./l2fwd -c 0xc -n 4 -- -q 1 -p 0xc

b. SlotA(CPU1) -> CPU0
#>./l2fwd -c 0xc0 -n 4 -- -q 1 -p 0xc0

Results:
use frame size 128 bytes
CPU Affinity

Slot A (CPU1)

Slot B (CPU0)

CPU0

15.9%

96.49%

CPU1

90.88%

24.78%

本信件可能包含瑞祺電通機密資訊，非指定之收件者，請勿使用或揭露本信件內容，並請銷毀此信件。 This email may contain confidential information. Please do not use or disclose it in any way and delete it if you are not the intended recipient.

Many DPDK API's have NUMA socket as one of the parameters. In order to get good performance it is up to the application to be NUMA aware and use socket local resources.

One example we do is to have a packet mbuf pool per socket, and assign each device to the correct pool. Also, you may want to choose which lcore's to assign to which function based on socket locality. For example threads that are polling receiver should be on same socket as that NIC.

Remember the example applications are demo toys, and don't do all the things a real application would need to do.

QPI bandwidth is definitely large enough, but it seems that QPI is only responsible for the communication between separate CPU chips. What you need to do is actually accessing the memory on the other part, probably not even hit the bandwidth. The latency can be caused by a lot of facts during a NUMA operation.

/Bob

------------------ 原始邮件 ------------------
发件人: "Zachary";<[email protected]>;
发送时间: 2013年9月2日(星期一) 中午11:22
收件人: "dev"<[email protected]>;
抄送: " "Yannic.Chou (周哲正) : 6808" <[email protected]>; "Alan Yu 俞亦偉 : 6632""<[email protected]>; 主题: [dpdk-dev] DPDK & QPI performance issue in Romley platform.

Hi~

I have a question about DPDK & QPI performance issue in Romley platform. Recently, I use DPDK example, l2fwd, to test DPDK's performance in my Romley platform. When I try to do the test, crossing used CPU, I find the performance dramatically decrease. Is it true? Or any method can prove the phenomenon?

In my opinion, there should be no this kind of issue here due to QPI have enough bandwidth to deal the kinds of case. Thus, I am so amaze in our results and can not explain it.
Could someone can help me to solve this problem.

Thank a lot!

My testing environment describe as below:

Platform: Romley
CPU: E5-2643 * 2
RAM: Transcend 8GB PC3-1600 DDR3 * 8
OS: Fedora core 14
DPDK: v1.3.1r2, example/l2fwd
Slot setting:
SlotA is controled by CPU1 directly. SlotB is controled by CPU0 directly.

DPDK pre-setting:
a. BIOS setting:
HT=disable
b. Kernel paramaters

     isolcpus=2,3,6,7
     default_hugepagesz=1024M
     hugepagesz=1024M
     hugepages=16

c. OS setting:
service avahi-daemon stop
service NetworkManager stop
service iptables stop
service acpid stop
selinux disable

Example program Command:
a. SlotB(CPU0) -> CPU1

     #>./l2fwd -c 0xc -n 4 -- -q 1 -p 0xc

b. SlotA(CPU1) -> CPU0

     #>./l2fwd -c 0xc0 -n 4 -- -q 1 -p 0xc0 

Results:
use frame size 128 bytes

CPU Affinity

Slot A (CPU1)

Slot B (CPU0)

CPU0

15.9%

96.49%

CPU1

90.88%

24.78%

本信件可能包含瑞祺電通機密資訊，非指定之收件者，請勿使用或揭露本信件內容，並請銷毀此信件。 This email may contain confidential information. Please do not use or disclose it in any way and delete it if you are not the intended recipient.