Re: net_sched: precision problem of TBF/HTB

[John Fastabend <john.fastabend@gmail.com>]

On 07/16/2013 06:41 PM, Ding Tianhong wrote:
> On 2013/7/16 12:50, Eric Dumazet wrote:
>> On Tue, 2013-07-16 at 12:12 +0800, Yang Yingliang wrote:
>>> Hi, Eric
>>> Commit 1def9238d4aa2 (net_sched: more precise pkt_len computation)
>>> makes more precise transfer bytes by taking account of headers in
>>> qdisc_skb_cb(skb)->pkt_len, but this introduces a problem with
>>> calculating bandwidth in userland.
>>>
>> This changed nothing to userland.
>>
>>> When calculating bandwidth in userland, it's not include headers'
>>> bytes. From the user's perspective, it's not a correct bandwidth.
>>>
>>
>>
>>> Shall we need take account of headers in qdisc_skb_cb(skb)->pkt_len or
>>> just skb->len?
>>>
>> These values are not accessible from userland, unless you capture
>> packets with a sniffer (tcpdump or something like that)
>>
>>> Example:
>>> tc qdisc add dev $DEV root handle 1: tbf latency 50ms burst 500kB rate
>>> 500mbit mtu 64k
>>>
>>> iperf -c host -t 30 -i 10
>>> [ ID] Interval       Transfer     Bandwidth
>>> [  3]  0.0-10.0 sec   571 MBytes   479 Mbits/sec
>>> [  3] 10.0-20.0 sec   570 MBytes   478 Mbits/sec
>>> [  3] 20.0-30.0 sec   570 MBytes   478 Mbits/sec
>>> [  3]  0.0-30.0 sec  1.67 GBytes   478 Mbits/sec
>>
>>
>> iperf only measures the amount of payload, and probably do not care of
>> headers.
>>
>> You cannot accurately measure bandwidth from userland, unless making
>> assumptions (or getting them from the stack) on header sizes (IP + TCP),
>> MSS value, and retransmits.
>>
>>
>
> ok, for the further, can you give me some advise for how calculate the bandwidth for net link.
> for example,what kind of tools is better.
>
>>

Vendor specific but it looks reasonably common to count tx and rx bytes
in ethtool stats. Take a few samples with 'ethtool -S' and do the
math. This will give you what the driver/hardware actually xmits and
receives.

.John

-- 
John Fastabend         Intel Corporation