[Devel] RE: Re: [HELP] for using class group classifier network controller

Chei-Yol Kim gauri at etri.re.kr
Wed Jun 17 03:03:04 PDT 2009


First, I thank you very much for your kind reply.

>Which is your problem? 
> 
>  a) network packet can't be classified into the class which is set to net_cls.classid 
> 
>  b) network packet can be classified into the class which is set to net_cls.classid, 
>     but its bandwidth limit doesn't work correctly. 
> 
>If your problem is b), unfortunately I can't help you. 
>Because I am a begginer of cls_cgroup, and I haven't tested enough. 
>(But I think I might be facing same problem.) 
> 
>If your problem is a), you should check 'tc -s filter show'. 
>You can see the result of classifing by cls_cgroup. 
>I can classified to the class which is set to root's net_cls.classid. 
> 
>Environment: 
>  kernel: 2.6.30 
>  tc: iproute2-2.6.29-1 with tc/f_cgroup.c patch in latest git. 
>      (I think it's same, probably) 
> 
>How to use: 
>  1) set to qdisc, class, filter 
> 
>     # tc qdisc add dev bond0 root handle 1: htb default 30 
> 
>     # tc class add dev bond0 parent 1:0 classid 1:10 htb rate 1mbit 
>     # tc class add dev bond0 parent 1:0 classid 1:20 htb rate 2mbit 
>     # tc class add dev bond0 parent 1:0 classid 1:30 htb rate 3mbit 
> 
>     # tc filter add dev bond0 protocol ip parent 1: prio 1 handle 0x1 cgroup 
> 
>     # tc -s class show dev bond0 
>     class htb 1:10 root prio 0 rate 1000Kbit ceil 1000Kbit burst 1724b cburst 1724b 
>      Sent 0 bytes 0 pkt (dropped 0, overlimits 0 requeues 0)                         <- packet count is 0 
>      rate 0bit 0pps backlog 0b 0p requeues 0 
>      lended: 0 borrowed: 0 giants: 0 
>      tokens: 13476 ctokens: 13476 
> 
>     class htb 1:20 root prio 0 rate 2000Kbit ceil 2000Kbit burst 1849b cburst 1849b 
>      Sent 0 bytes 0 pkt (dropped 0, overlimits 0 requeues 0) 
>      rate 0bit 0pps backlog 0b 0p requeues 0 
>      lended: 0 borrowed: 0 giants: 0 
>      tokens: 7226 ctokens: 7226 
> 
>     class htb 1:30 root prio 0 rate 3000Kbit ceil 3000Kbit burst 1974b cburst 1974b 
>      Sent 25804 bytes 214 pkt (dropped 0, overlimits 0 requeues 0) 
>      rate 3992bit 4pps backlog 0b 0p requeues 0 
>      lended: 214 borrowed: 0 giants: 0 
>      tokens: 4851 ctokens: 4851 
> 
>  2) mount cls_cgroup and set root's net_cls.classid 
> 
>     # mount -t cgroup -onet_cls none /cgroup/net/ 
>     # cd /cgroup/net/ 
>     # ls 
>     net_cls.classid  notify_on_release  release_agent  tasks 
> 
>     # echo 0x10010 > net_cls.classid                 # I want to classify 1:10 
>     # cat net_cls.classid 
>     65552                                            # 65552 = 65536 + 16 = 0x10010 
> 
>  3) using network 
> 
>  4) check the packet which is classified to the class which is set to net_cls.classid. 
> 
>     # /home/usui/src/iproute2-2.6.29-1/tc/tc -s class show dev bond0 
>     class htb 1:10 root prio 0 rate 1000Kbit ceil 1000Kbit burst 1724b cburst 1724b 
>      Sent 11238 bytes 95 pkt (dropped 0, overlimits 0 requeues 0)                    <- increase packet count 
>      rate 2944bit 3pps backlog 0b 0p requeues 0 
>      lended: 95 borrowed: 0 giants: 0 
>      tokens: 12601 ctokens: 12601

I've done the same experiment above your configuration varing each rate of classes. 
I set the rate as 10mbit, 20mbit, 30mbit for easy measuring.

when I set the classid to 0x10010, class 1:10's packet count was increased. 
But, according to your configuration, every packet has to flowed to 1:10 class. 
Of course I know that the net_cls can't classify all packet because of the getting classid from the current task. 
Even though we consider it, Many packets are flowed to 1:30 class. 
therefore, the rate of the socket showed about 30mbps. This is the 1:30 class's rate.
I think it's because that 1:30 class is the default class. 
When I changed the default to 20, the rate was varied to 20mbps, class 1:20's rate.

To measure this rate, I used the iperf and scp command. Two results are the same.
How about your experiment's result? 
Would you check the accurate rate of the network using applications like a iperf?

By the results of my experiments, net_cls couldn't control the network rate of the group. 
What is your opinion?

> 
>     class htb 1:20 root prio 0 rate 2000Kbit ceil 2000Kbit burst 1849b cburst 1849b 
>      Sent 0 bytes 0 pkt (dropped 0, overlimits 0 requeues 0) 
>      rate 0bit 0pps backlog 0b 0p requeues 0 
>      lended: 0 borrowed: 0 giants: 0 
>      tokens: 7226 ctokens: 7226 
> 
>     class htb 1:30 root prio 0 rate 3000Kbit ceil 3000Kbit burst 1974b cburst 1974b 
>      Sent 46092 bytes 342 pkt (dropped 0, overlimits 0 requeues 0) 
>      rate 320bit 0pps backlog 0b 0p requeues 0 
>      lended: 342 borrowed: 0 giants: 0 
>      tokens: 5017 ctokens: 5017

- Chei-yol Kim


_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers




More information about the Devel mailing list