[dpdk-users] Using 1Gbps NIC to send packets, but only get 100Mbps speed

[dpdk-users] Using 1Gbps NIC to send packets, but only get 100Mbps speed

[xiaozhe shao]

Hi all,

	I’m sending packet by rte_eth_tx_burst() function. With 1Gbps NIC, I just get 100Mbps speed.
	I use the following code to send packets (In order to send packets, I keep calling rte_eth_tx_burst).

for (i=0; i < BURST_SIZE;){
	i = i + rte_eth_tx_burst(port, queue, buffer + i, BURST_SIZE - i);
}

	My NICs are “Intel Corporation 82572EI Gigabit Ethernet Controller (Copper) (rev 06)” and “Intel Corporation 82576 Gigabit Ethernet Connection (rev 01)”. When I get the link state through DPDK, the link_speed is 1000. 

	I try testpmd with the instructions of “http://dpdk.org/doc/quick-start <http://dpdk.org/doc/quick-start>”. It works well. No dropped packets.

	When I send packets with 60 bits length, the rate is only about 140000 pps. When I send packets with 1500 bits length, the rate is about 8000 pps. I only get 100Mbps sending speed. 

	My OS is “Linux 3.2.0-4-amd64 #1 SMP Debian 3.2.81-1 x86_64 GNU/Linux”. I tried DPDK-2.2.0 and DPDK-16.07. The results are same.

	I also found a strange thing. If I send the packet that has the destination MAC same with the MAC address of the NIC that sends the packet, I can get 1.4Mpps with the 60-bit-length packets.

	I profile the sending program with gperf. The “rte_delay_us” consumes more than 90% running time. But I don’t know the caller of “rte_delay_us”. In the output of gperf, “rte_delay_us” is “spontaneous”.

	I don’t know why I cannot get 1Gbps send speed. Did anyone meet the similar issue? How to address it?

Best regards,
-xiaozhe

Re: [dpdk-users] Using 1Gbps NIC to send packets, but only get 100Mbps speed

[xiaozhe shao]

Hi all,

	I found that the destination MAC address is the MAC address of an active NIC (managed by Linux OS), I can get roughly line rate (1Gbps). If the destination NIC is bind to the DPDK driver, I can only get 100Mbps. All NICs are Gigabit-NIC. I use D-Link DGS-1024D Gigabit switch for the NIC connection.
	
	It is not the PAUSE frame. Because, I already disable the flow control on NICs. Is there any other mechanism that limits the packet sending speed? I googled for long time, but didn’t find any possible reason.

	Does anyone have any suggestion? Thanks.

Best regards,
-xiaozhe

> On Sep 20, 2016, at 11:09 AM, xiaozhe shao <sxz1069@gmail.com> wrote:
> 
> Hi all,
> 
> 	I’m sending packet by rte_eth_tx_burst() function. With 1Gbps NIC, I just get 100Mbps speed.
> 	I use the following code to send packets (In order to send packets, I keep calling rte_eth_tx_burst).
> 
> for (i=0; i < BURST_SIZE;){
> 	i = i + rte_eth_tx_burst(port, queue, buffer + i, BURST_SIZE - i);
> }
> 
> 	My NICs are “Intel Corporation 82572EI Gigabit Ethernet Controller (Copper) (rev 06)” and “Intel Corporation 82576 Gigabit Ethernet Connection (rev 01)”. When I get the link state through DPDK, the link_speed is 1000. 
> 
> 	I try testpmd with the instructions of “http://dpdk.org/doc/quick-start <http://dpdk.org/doc/quick-start>”. It works well. No dropped packets.
> 
> 	When I send packets with 60 bits length, the rate is only about 140000 pps. When I send packets with 1500 bits length, the rate is about 8000 pps. I only get 100Mbps sending speed. 
> 
> 	My OS is “Linux 3.2.0-4-amd64 #1 SMP Debian 3.2.81-1 x86_64 GNU/Linux”. I tried DPDK-2.2.0 and DPDK-16.07. The results are same.
> 
> 	I also found a strange thing. If I send the packet that has the destination MAC same with the MAC address of the NIC that sends the packet, I can get 1.4Mpps with the 60-bit-length packets.
> 
> 	I profile the sending program with gperf. The “rte_delay_us” consumes more than 90% running time. But I don’t know the caller of “rte_delay_us”. In the output of gperf, “rte_delay_us” is “spontaneous”.
> 
> 	I don’t know why I cannot get 1Gbps send speed. Did anyone meet the similar issue? How to address it?
> 
> Best regards,
> -xiaozhe