tc doesn't shape correctly -->> iproute2-latest + iptables 1.4.3.2 +linux-kernel 2.6.29.4 + imq support

tc doesn't shape correctly -->> iproute2-latest + iptables 1.4.3.2 +linux-kernel 2.6.29.4 + imq support

[Andrei Moraru]

Hello
I am running a Linux gateway, with details written in title.
The kernel is patched for imq support.
The gateway has 2 interfaces, eth0 --> ISP, , eth1 -->LAN.
Also there w virtual devices, imq0 atached to eth0 for download from
the world, , imq1 atached to eth1 for upload from the LAN.
I have a problem with traffic shaping:
1) the ceil value for the root class atached to imq0 is not respected;
currently the ceil value is 3200 kbits but the bandwitdh goes near
3700 kbits.
Some details of my system that might be helpful:
CONFIG_HZ=1000
net.ipv4.netfilter.ip_conntrack_max = 262140
The discipline for queueing is HFSC.
Best regards,
  matagou

Re: tc doesn't shape correctly -->> iproute2-latest + iptables 1.4.3.2 +linux-kernel 2.6.29.4 + imq support

[Karel Rericha]

2009/10/8 Andrei Moraru <matagou@gmail.com>:
> I have a problem with traffic shaping:
> 1) the ceil value for the root class atached to imq0 is not respected;
> currently the ceil value is 3200 kbits but the bandwitdh goes near
> 3700 kbits.

Same here,

I can confirm that similar setup (no imq) on Gentoo quadcore AMD64
2.6.29 HFSC root class overflows where HTB is ok. This situation
persist for about year and few last kernels and I havent found a
solution except some notices about HFSC in current linux incarnation
being buggy.

Our solution was return back to HTB :-)

If anyone knows why HFSC overflows, please post.

Karel

P.S.: Half year back we tested simple HFSC setup on bench with 3
classes (5, 10 and 15Mbits with root class 20Mbits) and when we
saturated all classes, total flow was 20Mbit for few seconds (5-30)
and then it overflowed to about 25Mbit, and sometimes returned back to
20Mbit after few minutes. VERY strange.

Re: tc doesn't shape correctly -->> iproute2-latest + iptables 1.4.3.2 +linux-kernel 2.6.29.4 + imq support

[Michal Soltys]

> 
> I can confirm that similar setup (no imq) on Gentoo quadcore AMD64
> 2.6.29 HFSC root class overflows where HTB is ok. This situation
> persist for about year and few last kernels and I havent found a
> solution except some notices about HFSC in current linux incarnation
> being buggy.
> 

Mind giving a pointer to such notices ?

> Our solution was return back to HTB :-)
> 
> If anyone knows why HFSC overflows, please post.
> 
> Karel
> 
> P.S.: Half year back we tested simple HFSC setup on bench with 3
> classes (5, 10 and 15Mbits with root class 20Mbits) and when we
> saturated all classes, total flow was 20Mbit for few seconds (5-30)
> and then it overflowed to about 25Mbit, and sometimes returned back to
> 20Mbit after few minutes. VERY strange.

What exactly was your setup and how did you test it ?

Re: tc doesn't shape correctly -->> iproute2-latest + iptables 1.4.3.2 +linux-kernel 2.6.29.4 + imq support

[Karel Rericha]

2009/10/13 Michal Soltys <soltys@ziu.info>:
>>
>> I can confirm that similar setup (no imq) on Gentoo quadcore AMD64
>> 2.6.29 HFSC root class overflows where HTB is ok. This situation
>> persist for about year and few last kernels and I havent found a
>> solution except some notices about HFSC in current linux incarnation
>> being buggy.
>>
>
> Mind giving a pointer to such notices ?

Sorry, I dont have links to such notices I read when I did tests, but
quick Google search for "hfsc overflows" is giving results like this:
http://patchwork.ozlabs.org/patch/27303/

>> P.S.: Half year back we tested simple HFSC setup on bench with 3
>> classes (5, 10 and 15Mbits with root class 20Mbits) and when we
>> saturated all classes, total flow was 20Mbit for few seconds (5-30)
>> and then it overflowed to about 25Mbit, and sometimes returned back to
>> 20Mbit after few minutes. VERY strange.
>
> What exactly was your setup and how did you test it ?
>

Not sure (long time ago), but I think gentoo-sources 2.6.27 AMD64 on
dual dualcore Opteron 2214, 2x Intel 82571EB, 2x Broadcom BCM5754,
tested without bonding (but in production setup bonding is used),
without packet offloading, no imq nor ifb used. HFSC overflowed but
HTB was ok.

Karel

Re: tc doesn't shape correctly -->> iproute2-latest + iptables 1.4.3.2 +linux-kernel 2.6.29.4 + imq support

[Michal Soltys]

Karel Rericha wrote:
> 
>>> P.S.: Half year back we tested simple HFSC setup on bench with 3
>>> classes (5, 10 and 15Mbits with root class 20Mbits) and when we
>>> saturated all classes, total flow was 20Mbit for few seconds (5-30)
>>> and then it overflowed to about 25Mbit, and sometimes returned back to
>>> 20Mbit after few minutes. VERY strange.
>> What exactly was your setup and how did you test it ?
>>
> 
> Not sure (long time ago), but I think gentoo-sources 2.6.27 AMD64 on
> dual dualcore Opteron 2214, 2x Intel 82571EB, 2x Broadcom BCM5754,
> tested without bonding (but in production setup bonding is used),
> without packet offloading, no imq nor ifb used. HFSC overflowed but
> HTB was ok.
> 
> Karel
> 

I meant more about your hfsc rules in that simple setup, not the hardware.