Pktgen detailed tutorial

There are a lot of articles on pktgen on the Internet, but it is always not comprehensive and meticulous. After reading it, I still won't write the pktgen test script. Therefore, this article will elaborate on pktgen in detail, after reading this article, you can write pktgen shell by yourself.

　　

#modprobe pktgen

Then you will see the following file in your computer's/proc/net/pktgen:

kpktgend_0  kpktgend_1  kpktgend_2  kpktgend_3  pgctrl

The number of kpktgen _ * is determined based on the number of your CPU. For example, if the number of sub-CPUs of my machine is 4, there are four such files.

Run cat/proc/net/pktgen/pgctrl to view the pktgen version and other information:

#cat /proc/net/pktgen/pgctrl
Packet Generator for packet performance testing. Version: 2.74

　　

　　1) interrupt affinity

When a thread (kpktgend_x) is used to send (accept) data through a port (ethx), we should bind the interrupt corresponding to this port to a certain CPU, this prevents CPU cache loss due to CPU changes (caused by the operating system. The procedure is as follows:

First, run the following command to view the ethx interrupt. Here we use eth3 as an example.

# cat /proc/interrupts | grep eth3

47: 45836 0 0 0 PCI-MSI-edge eth3-TxRx-0
48: 45836 0 0 0 PCI-MSI-edge eth3-TxRx-1
49: 45836 0 0 0 PCI-MSI-edge eth3-TxRx-2
50: 45836 0 0 0 PCI-MSI-edge eth3-TxRx-3
51: 0 0 0 0 PCI-MSI-edge eth3: lsc

From the above we can see that eth3 uses five interrupt numbers, 47-51. Here we will bind them to CPU0.

# echo 1 /proc/irq/47/smp_affinity# echo 1 /proc/irq/48/smp_affinity# echo 1 /proc/irq/49/smp_affinity# echo 1 /proc/irq/50/smp_affinity# echo 1 /proc/irq/51/smp_affinity

Of course, if you bind them separately, you can share the CPU burden during the big data stream, which may improve the maximum performance, but it is unclear, this should also be related to the multi-queue of the Intel Nic. Further experiments are needed to analyze the problem.

According to Robert's test results, correct interruption affinity binding improves the sending rate by 60%.

　　2) clone_skb: limits the memory allocation (clone_skb should only copy the structure of skb and not copy the data zone)

This part has not been understood yet. It should be related to the high-speed cache of the cpu. It should be put first. It does not affect the subsequent experiments.

　　3) latency: reduces the transmission rate

After a packet is sent, the unit of delay is ns level, so as to reduce the sending rate. If the delay time is greater than the sending time, you can use this function to send a specific number of data packets per second, that is, the frequency of anset to send data packets.

 

　　

 

　1) The experiment topology is as follows:

 

1+-----------+           +-----------+2| Host1     |           |  Host2     |3|       eth1 +-----------+nf0       |4|           |           |           |5|           |           |           |67+-----------+           +-----------+

 

First, let's take a look at the next simple code, which gives you an intuitive understanding of the shell control of pktgen. You can try to see this code first, or skip it directly. First, let's take a look at the following content, then I will check the code again.

 [[ `lsmod | grep pktgen` ==  [[  == =550= > =`cat  | fgrep  [  =  |> =/proc/net/pktgen/echo =/proc/net/pktgen/=/proc/net/pktgen/

　

　　2) All pktgen commands are explained as follows:

PktgenCommand
PktgenControl Command
Start	All threads start sending
Stop	Stop
Thread control command
Add_device	Add a port to a thread
Rem_device_all	Delete all ports bound to a thread
Max_before_softirq	Run do_softirq ()
PORT command
Debug	Debugging
Clone_skb	How many copies are performed on each skb? 0 indicates no replication. Dos and other tests must be zero
Clear_counters	Clear the counter. The general program automatically clears the counter.
Pkt_size	The size of the link package (CRC value)
Min_pkt_size	Minimum packet Value
Max_pkt_size	Maximum Value
Flags	Number of packages?
Count	Number of sent data packets. 0 indicates that data packets are sent all the time.
Delay	Latency between two data packets
Dst	Destination IP address
Dst_min	Minimum destination IP address
Dst_max	Maximum Value
Src_min	Minimum source IP address
Src_max	Maximum Value
Dst6	Destination IPv6 address
Src6	Source IPv6 address
Dstmac	Target mac
Srcmac	Source mac
Src_mac_count	Number of source mac, which starts from the mac set by srcmac.
Dst_mac_count	Same as above
Udp_src_min	Minimum source udp port number
Udp_src_max	Maximum source udp port number
Udp_dst_min	Minimum destination udp port number
Udp_dst_max	Maximum destination udp port number
Flows	Number of concurrent streams
Flowlen	Stream Length
Flags
IPSRC_RND IPDST_RND TXSIZE_RND UDPSRC_RND UDPDST_RND MACSRC_RND MACDST_RND	PSRC_RND source IP address random sending

　　

 

 > =`cat  | fgrep  [  =  |> 

 

Pktgen. conf-1: When the CPU system binds eth1 eth2 to Thread 0, part of the code is as follows

PGDEV=/proc/net/pktgen/

Pktgen. conf-2: eth1 uses Thread 0 eth2 uses thread 1 to send, part of the code is as follows

PGDEV=/proc/net/pktgen/=/proc/net/pktgen/

Pktgen. conf-3: In a dual CPU system, eth1 eth2 uses Thread 0 to send, part of the code is as follows

PGDEV=/proc/net/pktgen/=/proc/net/pktgen/

Pktgen. conf-4: the script for Dos testing, pay attention to clone_skb 0, part of the code is as follows

PGDEV=/proc/net/pktgen/

Pktgen. conf-5: the script for routing flow testing, also pay attention to clone_skb 0

PGDEV=/proc/net/pktgen/pgset 

Finally, paste a complicated script:

PGDEV=/proc/net/pktgen/=/proc/net/pktgen/ i  2 3 4 5 6 7 8 9 10 11=/proc/net/pktgen/eth=/proc/net/pktgen/-2 /proc/net/pktgen/eth*

Related links:

Linux package tools pktgen tutorial @ cangqiong fantasy