LARTC - May 2005 - Question: traffic shaping (tc-htb)

Hello *!
My trafficshaping (tc-htb) drops packets very early, at least i suspect
this. It drops about 30% fo the packets. The traffic-generating appication
is running locally on the shaping host. I think i can lower this rate by
increasing the packet-buffer, because the locally application will slow down
with increasing buffer (tcp/ip). But i can´t find any options for that, i
cycled the manual several times. Is there a solution?




see output of "tc -s -d class show dev ppp0" (main class):
    class htb 1:10 parent 1:1 prio 1 quantum 1800 rate 144000bit
        ceil 480000bit burst 1779b/8 mpu 0b overhead 0b cburst 2199b/8
        mpu 0b overhead 0b level 0
    Sent 57567157 bytes 40609 pkts (dropped 16033, overlimits 0)
    rate 460328bit 40pps
    lended: 12845 borrowed: 27764 giants: 0
    tokens: -74623 ctokens: -28826

This class is the only one with packet drops. Does tc use the kernels
network packet queue?


Great thanks in advance
Alvo

John Smith wrote:
> Hello *!
> My trafficshaping (tc-htb) drops packets very early, at least i suspect
> this. It drops about 30% fo the packets. The traffic-generating appication
> is running locally on the shaping host. I think i can lower this rate by
> increasing the packet-buffer, because the locally application will slow
down
> with increasing buffer (tcp/ip). But i can´t find any options for that, i
> cycled the manual several times. Is there a solution?
> 
> 
> 
> 
> see output of "tc -s -d class show dev ppp0" (main class):
>     class htb 1:10 parent 1:1 prio 1 quantum 1800 rate 144000bit
>         ceil 480000bit burst 1779b/8 mpu 0b overhead 0b cburst 2199b/8
>         mpu 0b overhead 0b level 0
>     Sent 57567157 bytes 40609 pkts (dropped 16033, overlimits 0)
>     rate 460328bit 40pps
>     lended: 12845 borrowed: 27764 giants: 0
>     tokens: -74623 ctokens: -28826
> 
> This class is the only one with packet drops. Does tc use the kernels
> network packet queue?
HTB uses the txqueuelength of the interface if you don''t add a queue to
the leaf class. For my ppp0 that''s 3 which is a bit short - even so I 
just did a test and only got 10% loss so maybe your generator app/kernel 
version of tcp is a bit over aggressive. I did 1 tcp stream with netperf 
on 2.6.12-rc1.

So either add a queue to the leaf and specify a length or before you 
start htb do ifconfig ppp0 txqueuelen 30 or whatever.


Andy.

Hello!
tc-htb is great because its easy, but you cannot specify an mpu. Thats bad,
because with the mpu you can specify the physical characteristics of the
underlying connection. I have a braodband dsl-connection. I have the problem
that the proportion of small packets to big packets changes a lot. So
without specifing the mpu the connection is not to full capacity (big
packets) or the buffer of the modem will be filled (small packets).

Can i append an tc-tbf qdisc (qdisc 30:) to an tc-htb class with specifiying
an mpu? I already did so for increasing the queue for big packets (qdisc
20:), but when i tried this for small packets i could not direct data with
an tc filter to it, it alway go to qdisc 20:.

tc tree Configuration:
    root qdisc htb
        class 1   htb
            class 1:10   htb
                qdisc 20: tbf
            class 1:11 htb
                qdisc 30: tbf

tc filter:
    tc filter add dev $dev protocol ip handle 10 \
        fw flowid 20:
    tc filter add dev $dev protocol ip handle 11 \
        fw flowid 30:


I don´t know how those queue work, if you specify the mpu in 20: or 30: will
the parent classes work with the calculated size of the packets? Anyway,
traffic was not directed in 30: :(

any suggestions?
Great thanks in advance
Alvo

John Smith wrote:
> Hello!
> tc-htb is great because its easy, but you cannot specify an mpu.
You can specify mpu and overhead with htb  - well you can with the 
recent 2.6 I use.

  Thats bad,
> because with the mpu you can specify the physical characteristics of the
> underlying connection. I have a braodband dsl-connection. I have the
problem
> that the proportion of small packets to big packets changes a lot. So
> without specifing the mpu the connection is not to full capacity (big
> packets) or the buffer of the modem will be filled (small packets).
If you can find your overhead for your dsl type then you can patch htb 
and tc to do it perfectly.

Have a look at the thesis on http://www.adsl-optimizer.dk/ there is a 
section that shows different overheads.

Andy.