Re: [Bloat] Initial tests with BBR in kernel 4.9

[Neal Cardwell <ncardwell@google.com>]

[-- Attachment #1: Type: text/plain, Size: 4592 bytes --]

On Wed, Jan 25, 2017 at 5:38 PM, Hans-Kristian Bakke <hkbakke@gmail.com>
wrote:

> Actually.. the 1-4 mbit/s results with fq sporadically appears again as I
> keep testing but it is most likely caused by all the unknowns between me an
> my testserver. But still, changing to pfifo_qdisc seems to normalize the
> throughput again with BBR, could this be one of those times where BBR and
> pacing actually is getting hurt for playing nice in some very variable
> bottleneck on the way?
>

Possibly. Would you be able to take a tcpdump trace of each trial (headers
only would be ideal), and post on a web site somewhere a pcap trace for one
of the slow trials?

For example:

   tcpdump -n -w /tmp/out.pcap -s 120 -i eth0 -c 1000000 &

thanks,
neal



>
> On 25 January 2017 at 23:01, Neal Cardwell <ncardwell@google.com> wrote:
>
>> On Wed, Jan 25, 2017 at 3:54 PM, Hans-Kristian Bakke <hkbakke@gmail.com>
>> wrote:
>>
>>> Hi
>>>
>>> Kernel 4.9 finally landed in Debian testing so I could finally test BBR
>>> in a real life environment that I have struggled with getting any kind of
>>> performance out of.
>>>
>>> The challenge at hand is UDP based OpenVPN through europe at around 35
>>> ms rtt to my VPN-provider with plenty of available bandwith available in
>>> both ends and everything completely unknown in between. After tuning the
>>> UDP-buffers up to make room for my 500 mbit/s symmetrical bandwith at 35 ms
>>> the download part seemed to work nicely at an unreliable 150 to 300 mbit/s,
>>> while the upload was stuck at 30 to 60 mbit/s.
>>>
>>> Just by activating BBR the bandwith instantly shot up to around 150
>>> mbit/s using a fat tcp test to a public iperf3 server located near my VPN
>>> exit point in the Netherlands. Replace BBR with qubic again and the
>>> performance is once again all over the place ranging from very bad to bad,
>>> but never better than 1/3 of BBRs "steady state". In other words "instant
>>> WIN!"
>>>
>>
>> Glad to hear it. Thanks for the test report!
>>
>>
>>> However, seeing the requirement of fq and pacing for BBR and noticing
>>> that I am running pfifo_fast within a VM with virtio NIC on a Proxmox VE
>>> host with fq_codel on all physical interfaces, I was surprised to see that
>>> it worked so well.
>>> I then replaced pfifo_fast with fq and the performance went right down
>>> to only 1-4 mbit/s from around 150 mbit/s. Removing the fq again regained
>>> the performance at once.
>>>
>>> I have got some questions to you guys that know a lot more than me about
>>> these things:
>>>
>> 1. Do fq (and fq_codel) even work reliably in a VM? What is the best
>>> choice for default qdisc to use in a VM in general?
>>>
>>
>> Eric covered this one. We are not aware of specific issues with fq in VM
>> environments. And  we have tested that fq works sufficiently well on Google
>> Cloud VMs.
>>
>>
>>> 2. Why do BBR immediately "fix" all my issues with upload through that
>>> "unreliable" big BDP link with pfifo_fast when fq pacing is a requirement?
>>>
>>
>> For BBR, pacing is part of the design in order to make BBR more "gentle"
>> in terms of the rate at which it sends, in order to put less pressure on
>> buffers and keep packet loss lower. This is particularly important when a
>> BBR flow is restarting from idle. In this case BBR starts with a full cwnd,
>> and it counts on pacing to pace out the packets at the estimated bandwidth,
>> so that the queue can stay relatively short and yet the pipe can be filled
>> immediately.
>>
>> Running BBR without pacing makes BBR more aggressive, particularly in
>> restarting from idle, but also in the steady state, where BBR tries to use
>> pacing to keep the queue short.
>>
>> For bulk transfer tests with one flow, running BBR without pacing will
>> likely cause higher queues and loss rates at the bottleneck, which may
>> negatively impact other traffic sharing that bottleneck.
>>
>>
>>> 3. Could fq_codel on the physical host be the reason that it still works?
>>>
>>
>> Nope, fq_codel does not implement pacing.
>>
>>
>>> 4. Do BBR _only_ work with fq pacing or could fq_codel be used as a
>>> replacement?
>>>
>>
>> Nope, BBR needs pacing to work correctly, and currently fq is the only
>> Linux qdisc that implements pacing.
>>
>>
>>> 5. Is BBR perhaps modified to do the right thing without having to
>>> change the qdisc in the current kernel 4.9?
>>>
>>
>> Nope. Linux 4.9 contains the initial public release of BBR from September
>> 2016. And there have been no code changes since then (just expanded
>> comments).
>>
>> Thanks for the test report!
>>
>> neal
>>
>>
>

[-- Attachment #2: Type: text/html, Size: 7910 bytes --]