Re: TCP funny-ness when over-driving a 1Gbps link.

[Ben Greear <greearb@candelatech.com>]

On 05/19/2011 05:46 PM, Rick Jones wrote:
> On Thu, 2011-05-19 at 17:37 -0700, Ben Greear wrote:
>> On 05/19/2011 05:24 PM, Rick Jones wrote:
>>>>>> [root@i7-965-1 igb]# netstat -an|grep tcp|grep 8.1.1
>>>>>> tcp        0      0 8.1.1.1:33038               0.0.0.0:*                   LISTEN
>>>>>> tcp        0      0 8.1.1.1:33040               0.0.0.0:*                   LISTEN
>>>>>> tcp        0      0 8.1.1.1:33042               0.0.0.0:*                   LISTEN
>>>>>> tcp        0 9328612 8.1.1.2:33039               8.1.1.1:33040               ESTABLISHED
>>>>>> tcp        0 17083176 8.1.1.1:33038               8.1.1.2:33037               ESTABLISHED
>>>>>> tcp        0 9437340 8.1.1.2:33037               8.1.1.1:33038               ESTABLISHED
>>>>>> tcp        0 17024620 8.1.1.1:33040               8.1.1.2:33039               ESTABLISHED
>>>>>> tcp        0 19557040 8.1.1.1:33042               8.1.1.2:33041               ESTABLISHED
>>>>>> tcp        0 9416600 8.1.1.2:33041               8.1.1.1:33042               ESTABLISHED
>>>>>
>>>>> I take it your system has higher values for the tcp_wmem value:
>>>>>
>>>>> net.ipv4.tcp_wmem = 4096 16384 4194304
>>>>
>>>> Yes:
>>>> [root@i7-965-1 igb]# cat /proc/sys/net/ipv4/tcp_wmem
>>>> 4096	16384	50000000
>>>
>>> Why?!?  Are you trying to get link-rate to Mars or something?  (I assume
>>> tcp_rmem is similarly set...)  If you are indeed doing one 1 GbE, and no
>>> more than 100ms then the default (?) of 4194304 should have been more
>>> than sufficient.
>>
>> Well, we occasionally do tests over emulated links that have several
>> seconds of delay and may be running multiple Gbps.  Either way,
>> I'd hope that offering extra RAM to a subsystem wouldn't cause it
>> to go nuts.
>
> It has been my experience that the autotuning tends to grow things
> beyond the bandwidthXdelay product.

Seems a likely culprit, or somehow it's not detecting round-trip-time
correctly, or maybe the timestamp is calculated when the pkt goes into
the send queue, and not when it's actually sent to the NIC?

>
> As for several seconds of delay and multiple Gbps - unless you are
> shooting the Moon, sounds like bufferbloat?-)

We try to test our stuff in all sorts of strange cases.  Maybe
some users really are emulating lunar traffic, or even beyond.
We also can emulate buffer bloat..but in this particular case,
real round-trip time is about 1-2ms, so if the socket is queuing up
a second's worth of bytes on the xmit buffer, then it's not
the network's fault...it's the sender.

>> Assuming this isn't some magical 1Gbps issue, you
>> could probably hit the same problem with a wifi link and
>> default tcp_wmem settings...
>
> Do you also increase tx queue's for the NIC(s)?

No, they are at the default (1000, I think).  That's only
a few ms at 1Gbps speed, so the problem is mostly higher
in the stack.

Thanks,
Ben

>
> rick


-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com