Re: excessive TCP dulplicate acks revisted

On Nov 10, 2007, at 10:28 AM, Andre Oppermann wrote:

> Gregory Wright wrote:
>> (Note: long message)
>> Hi,
>> The tcp duplicate ACK attack is back.
>> Last March, there was a thread on duplicate TCP acks in -CURRENT.
>> I have been able to reproduce the problem on 7.0-BETA2 (amd64)
>> and have some new information that might help locate the bug.
>> Background:  I first noticed problems with tcp connections dropping
>> when Bacula was running on our backup server.  The hardware was
>> a dual Opteron 244, 2 GB RAM, running FreeBSD 6.2-RELEASE-p2.
>> The ethernet NIC was a bge.
>> We went through an extensive process to rule out hardware and cabling
>> problems:  the server hardware was replaced with a single Opteron 270
>> (dual core) on a Tyan S2882-D motherboard.  The memory was replaced
>> as well.  The NICs are still bge.  This machine is "hardtack".
>
> This may be a TSO bug in the bge hardware.  Please repeat your  
> tests in
> the original setup with TSO disabled on the bge interfaces  
> ("ifconfig bgeX -tso").
> Second if you've got another network card with something else than bge
> please put it into the same box and run the tests as well.
>
> -- 
> Andre
>

Hi Andre,

I turned off TSO on the bge interface and still had the low  
throughput.  A packet
log showed that there were still bursts of duplicate ACKs.  Results  
of netperf
from my PowerBook:

crossroads-able> netperf -H hardtack
TCP STREAM TEST from localhost (0.0.0.0) port 0 AF_INET to hardtack. 
18clay.com (192.168.2.61) port 0 AF_INET
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

65536 262144 262144    10.02      28.37
crossroads-able>

The Tyan S2882-D motherboard has two bge interfaces and an fxp as well.
Switching to the fxp interface, I get much better throughput:

crossroads-able> netperf -H hardtack
TCP STREAM TEST from localhost (0.0.0.0) port 0 AF_INET to hardtack. 
18clay.com (192.168.2.61) port 0 AF_INET
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

65536 262144 262144    10.02      93.70
crossroads-able>

and the packet logs show no duplicate ACKs.

I will run some longer tests to the fxp interface to verify
that there are no problems.

Greg