Skip to contentSkip to conversation

HTB - problem with one thread

LARTC Ciprian Nica 2 participants 3 messages Aug 28, 2002 – Aug 29, 2002

lore
Share & tools
lore
#0 lore raw

Hello,

Besides a very good functionality, I have a small, yet stressing, problem.

Let's say I have a bandwidth of N bits.

I have X clients, everyone with his fixed bandwidth.

I create htb qdsics for each client. Each leaf has sfq queueing discipline.

A sample config would be like this :

tc class add dev eth0 parent 1:2 classid 1:346 htb rate 32Kbit ceil 32Kbit
burst 4k cburst 2k prio 4
tc qdisc add dev eth0 parent 1:346 handle 346 sfq perturb 10

AND

tc class add dev eth0 parent 1:2 classid 1:161 htb rate 256Kbit ceil
256Kbit burst 32k cburst 2k prio 4
tc qdisc add dev eth0 parent 1:161 handle 161 sfq perturb 10

The filters are like :
tc filter add dev eth0 parent 1:0 protocol ip prio 100 u32 match ip dst 1.2.3.128/25 classid 1:346 tc filter add dev eth0 parent 1:0 protocol ip prio 100 u32 match ip dst 1.2.3.4/30 classid 1:161

The problem is that clients can have their bandwidths at maximum only if they use a download accelerating software like DAP with multiple symultanous threads. Using only one thread, a single download can reach about 60% of the allocated bandwidth.

I have tested a download without any limits and it was working well, so there's no problem with the remote site.

I have searched thorugh messages posted on this mailing list but I didn't find a similar problem.

Maybe I need to set quantum manually. I have tested with r2q default value
(10) and also with 2 (raising quantum).

Another question would be: if quantum is bigger, the client can exceed his limit by that value ? Is that possible all the time or at bursty intervals ?

Thank you very much,
Ciprian Nica

_______________________________________________
LARTC mailing list / [email protected]
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

#1 lore raw
Quoted 41 lines · Ciprian Nica

Hello,

Besides a very good functionality, I have a small, yet stressing, problem.

Let's say I have a bandwidth of N bits.

I have X clients, everyone with his fixed bandwidth.

I create htb qdsics for each client. Each leaf has sfq queueing discipline.

A sample config would be like this :

tc class add dev eth0 parent 1:2 classid 1:346 htb rate 32Kbit ceil 32Kbit
burst 4k cburst 2k prio 4
tc qdisc add dev eth0 parent 1:346 handle 346 sfq perturb 10

AND

tc class add dev eth0 parent 1:2 classid 1:161 htb rate 256Kbit ceil
256Kbit burst 32k cburst 2k prio 4
tc qdisc add dev eth0 parent 1:161 handle 161 sfq perturb 10

The filters are like :
tc filter add dev eth0 parent 1:0 protocol ip prio 100 u32 match ip dst 1.2.3.128/25 classid 1:346 tc filter add dev eth0 parent 1:0 protocol ip prio 100 u32 match ip dst 1.2.3.4/30 classid 1:161

The problem is that clients can have their bandwidths at maximum only if they use a download accelerating software like DAP with multiple symultanous threads. Using only one thread, a single download can reach about 60% of the allocated bandwidth.

I have tested a download without any limits and it was working well, so there's no problem with the remote site.

I have searched thorugh messages posted on this mailing list but I didn't find a similar problem.

Quoted 6 lines

Maybe I need to set quantum manually. I have tested with r2q default value
(10) and also with 2 (raising quantum).

Another question would be: if quantum is bigger, the client can exceed his limit by that value ? Is that possible all the time or at bursty intervals ?

Quantum is used if 2 or more classes with the same parent are fighting for bandwidth. First, they are allowed to get the rate you gave them. For the remaining bandwidth, quantum is used. Each class may send quantum bytes. You have to make sure quantum > MTU (typical 1500 bytes for ethernet).

What if you use no sfq qdiscs?

Stef

Signature · 8 lines
[email protected]
 "Using Linux as bandwidth manager"
     http://www.docum.org/
     #lartc @ irc.oftc.net

_______________________________________________
LARTC mailing list / [email protected]
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/
#2 lore raw
In reply to Ciprian Nica
Quoted 5 lines · Stef Coene
Quoted 4 lines

The problem is that clients can have their bandwidths at maximum only if they use a download accelerating software like DAP with multiple symultanous threads. Using only one thread, a single download can reach about 60% of the allocated bandwidth.

Quoted 8 lines

Quantum is used if 2 or more classes with the same parent are fighting for bandwidth. First, they are allowed to get the rate you gave them. For the remaining bandwidth, quantum is used. Each class may send quantum bytes. You have to make sure quantum > MTU (typical 1500 bytes for ethernet).

What if you use no sfq qdiscs?

Stef

I have tried with pfifo queues and it's the same. I made many tests changing quantum, queueing for leaf but in no case I could reach the allocated bandwidth with a single download thread. The same limit set on cisco router with traffic-shape, worked fine.

Could there be a problem when there are many qdiscs ? Are there many calculations to be done, so the packets get delayed and the TCP transfer rate is lowered.

Ciprian

_______________________________________________
LARTC mailing list / [email protected]
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/