Skip to contentSkip to conversation

nesting htbs

LARTC Edward Smith 4 participants 6 messages Jun 9, 2005 – Jun 15, 2005

lore
Share & tools
lore
#0 lore raw

Hello all,
I am running a coop satellite link for my aviation company here in
Iraq. (silly blog www.stardotstar.org). I am running tc with htb with good success so far. I am working on improving it though and need some help. Currently I have just 4 classes, syn/ack/ping, webchat, http, and then other. We are really happy with how this has improved our ability to call home from our rooms and do video chat.
However, I would like to do a better job of making sure that each
IP is getting a fair share because it seems like sometimes one video or audio chat bullies another one into slowing down and one guy is having a great video and audio feed while someone elses audio only is suffering. I've seen some references to wrr and also to making a class for each IP. There doesn't seem to be much current documention on wrr, so I'm trying to set up nested htbs. Here are my questions:

1. Which makes more sense, to nest my 4 classes of traffic inside of a class for each IP, or to make a class for each IP in each of my 4 classes. I'm leaning towards the latter so that someones web traffic can't borrow from the interactive traffic classes.

2. I've done a test, and can't get any traffic into the nested classes. Here is my code:

#1:20 LOW DELAY--CHAT DATA
#includes the minimize delay FW TOS
tc class add dev ${UPDEV} parent 1:1 classid 1:20 htb rate 200kbit ceil ${UPCEIL}kbit burst 6k prio 1
tc filter add dev ${UPDEV} protocol ip parent 1: pref 20 u32\

       match ip tos 0x10 0xff flowid 1:20

tc filter add dev ${UPDEV} protocol ip parent 1: pref 21 handle 5 fw classid 1:20
tc filter add dev ${UPDEV} protocol ip parent 1: pref 21 handle 6 fw classid 1:20
tc qdisc add dev ${UPDEV} parent 1:20 handle 120: sfq perturb 10 nextclass 00
for clientip in cat /etc/ethers | awk '{ print $2 }'; do

       tc class add dev ${UPDEV} parent 1:20 classid 1:${nextclass}
htb rate  ${CLIENTRATE}kbit ceil ${CLIENTCEIL}kbit
       tc filter add dev ${UPDEV} protocol ip parent 1:20 prio 1 u32 \
               match ip src ${clientip} flowid 1:${nextclass}
       tc qdisc add dev ${UPDEV} parent 1:${nextclass} handle
${nextclass}: sfq perturb 10  #not sure if this is necessa
ry
       ((nextclass++))

done

I'm missing something, but not sure what. This code works as far as classes goes, just the filtering doesn't get traffic into the nested classes. I'm on 2.6.11 gentoo. Sorry about the wrapping, I'll look into setting nowrap on gmail.

Thanks!

Edward
_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#1 lore raw
Quoted 15 lines · Edward Smith

Hello all,
I am running a coop satellite link for my aviation company here in
Iraq. (silly blog www.stardotstar.org). I am running tc with htb with good success so far. I am working on improving it though and need some help. Currently I have just 4 classes, syn/ack/ping, webchat, http, and then other. We are really happy with how this has improved our ability to call home from our rooms and do video chat.
However, I would like to do a better job of making sure that each
IP is getting a fair share because it seems like sometimes one video or audio chat bullies another one into slowing down and one guy is having a great video and audio feed while someone elses audio only is suffering. I've seen some references to wrr and also to making a class for each IP. There doesn't seem to be much current documention on wrr, so I'm trying to set up nested htbs. Here are my questions:

Quoted 4 lines

1. Which makes more sense, to nest my 4 classes of traffic inside of a class for each IP, or to make a class for each IP in each of my 4 classes. I'm leaning towards the latter so that someones web traffic can't borrow from the interactive traffic classes.

I would do the latter also. I would have just one interactive class and give it a rate that is say 3/4 of the ceil, the bulk classes can still borrow the unused.

Quoted 10 lines

2. I've done a test, and can't get any traffic into the nested classes. Here is my code:

#1:20 LOW DELAY--CHAT DATA
#includes the minimize delay FW TOS
tc class add dev ${UPDEV} parent 1:1 classid 1:20 htb rate 200kbit ceil ${UPCEIL}kbit burst 6k prio 1
tc filter add dev ${UPDEV} protocol ip parent 1: pref 20 u32\

       match ip tos 0x10 0xff flowid 1:20
Quoted 8 lines

tc filter add dev ${UPDEV} protocol ip parent 1: pref 21 handle 5 fw classid 1:20
tc filter add dev ${UPDEV} protocol ip parent 1: pref 21 handle 6 fw classid 1:20

tc qdisc add dev ${UPDEV} parent 1:20 handle 120: sfq perturb 10 nextclass 00

You don't need this as it's not a leaf.

Quoted 2 lines

for clientip in cat /etc/ethers | awk '{ print $2 }'; do

If clientip is local because you are NATing than it won't work because traffic will have the real ip here.

To work around you could use marks. As you already use them for some things you may want to use --or-mark and u32 to match them eg.

iptables -A POSTROUTING -t mangle -p icmp -j MARK --set-mark 0x0100

and so on for traffic types using high byte then use low byte and --or-mark for addresses

iptables -A POSTROUTING -t mangle -s 192.168.0.1 -j MARK --or-mark 0x0001

Then filter top level with a mask like

tc filter add dev eth0 parent 1:0 protocol ip prio 1 u32 match mark 0x0100 0xff00 flowid 1:20

and leaf levels

tc filter add dev eth0 parent 1:20 protocol ip prio 1 u32 match mark 0x0001 0x00ff flowid 1:200

That assumes you really need iptables for marking traffic type - if you could use tc filters for that, then just use iptables for the addresses.

Quoted 7 lines
       tc class add dev ${UPDEV} parent 1:20 classid 1:${nextclass}
htb rate  ${CLIENTRATE}kbit ceil ${CLIENTCEIL}kbit
       tc filter add dev ${UPDEV} protocol ip parent 1:20 prio 1 u32 \
               match ip src ${clientip} flowid 1:${nextclass}
       tc qdisc add dev ${UPDEV} parent 1:${nextclass} handle

${nextclass}: sfq perturb 10 #not sure if this is necessa ry

I wouldn't put sfq on interactive - I would add a bfifo so I could set and play with the buffer lengths.

Andy.

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#2 lore raw
In reply to Edward Smith
Quoted 16 lines · On Fri, 10 Jun 2005 00:02:42 +0100 Andy Furniss <[email protected]>…
Quoted 15 lines · Edward Smith

Hello all,
I am running a coop satellite link for my aviation company here in
Iraq. (silly blog www.stardotstar.org). I am running tc with htb with good success so far. I am working on improving it though and need some help. Currently I have just 4 classes, syn/ack/ping, webchat, http, and then other. We are really happy with how this has improved our ability to call home from our rooms and do video chat.
However, I would like to do a better job of making sure that each
IP is getting a fair share because it seems like sometimes one video or audio chat bullies another one into slowing down and one guy is having a great video and audio feed while someone elses audio only is suffering. I've seen some references to wrr and also to making a class for each IP. There doesn't seem to be much current documention on wrr, so I'm trying to set up nested htbs. Here are my questions:

Quoted 23 lines
Quoted 4 lines

1. Which makes more sense, to nest my 4 classes of traffic inside of a class for each IP, or to make a class for each IP in each of my 4 classes. I'm leaning towards the latter so that someones web traffic can't borrow from the interactive traffic classes.

I would do the latter also. I would have just one interactive class and give it a rate that is say 3/4 of the ceil, the bulk classes can still borrow the unused.

Quoted 14 lines

2. I've done a test, and can't get any traffic into the nested classes. Here is my code:

#1:20 LOW DELAY--CHAT DATA
#includes the minimize delay FW TOS
tc class add dev ${UPDEV} parent 1:1 classid 1:20 htb rate 200kbit ceil ${UPCEIL}kbit burst 6k prio 1
tc filter add dev ${UPDEV} protocol ip parent 1: pref 20 u32\

       match ip tos 0x10 0xff flowid 1:20

tc filter add dev ${UPDEV} protocol ip parent 1: pref 21 handle 5 fw classid 1:20
tc filter add dev ${UPDEV} protocol ip parent 1: pref 21 handle 6 fw classid 1:20

so marking is thing that i need for limiting NATed uploads to internet ? when limiting downloads i do not need marking ?

am i right ?

Signature · 9 lines
*Dariusz 'tdi' Dwornikowski | Gentoo | admin at pozman.pl     |
*[JID]:[email protected]|[gg]:2266034|[IRC]:#gentoo-pl@freenode   |
*[MAIL]:[email protected]|[WWW]:www.tdi.pozman.pl                 | 
*Serwery,administracja,webapps - www.ProAdmin.com.pl          |
*Fingerprint:43E21CC46DAFD2F754E91547D59B39F56AAA4B5F         |
_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc
#4 lore raw
In reply to Edward Smith
Quoted 20 lines · Andy Furniss

If clientip is local because you are NATing than it won't work because traffic will have the real ip here.

To work around you could use marks. As you already use them for some things you may want to use --or-mark and u32 to match them eg.

iptables -A POSTROUTING -t mangle -p icmp -j MARK --set-mark 0x0100

and so on for traffic types using high byte then use low byte and --or-mark for addresses

iptables -A POSTROUTING -t mangle -s 192.168.0.1 -j MARK --or-mark 0x0001

Then filter top level with a mask like

tc filter add dev eth0 parent 1:0 protocol ip prio 1 u32 match mark 0x0100 0xff00 flowid 1:20

Quoted 7 lines

and leaf levels

tc filter add dev eth0 parent 1:20 protocol ip prio 1 u32 match mark 0x0001 0x00ff flowid 1:200

That assumes you really need iptables for marking traffic type - if you could use tc filters for that, then just use iptables for the addresses.

Something I've only just noticed from a comment in the code - htb can use mark without the need for lots of filters.

You only need one empty filter on the root (maybe you can still nest) like -

tc filter add dev eth0 parent 1:0 protocol ip prio 1 fw

and then if you arrange for your classes to be the same minor numbers as the marks it will behave like using classify.

You need to set the major number of your htb (1 in example above) in the top 16 bits of the mark.

There is also a netfilter pom-ng patch IPMARK that will set marks based on ipaddress.

Andy.
_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#5 lore raw
In reply to Edward Smith
Quoted 3 lines

I wouldn't put sfq on interactive - I would add a bfifo so I could set and play with the buffer lengths.

I agree. I think SFQ might reorder packets? It sometimes seems to cause some difficult to trace gremlins on my VoIP stuff, which might be due to packet re-ordering?

Best to stick with a bfifo I think

Ed W
_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc