[geeks] Help: Network issue has me stumped

Joshua Snyder josh at imagestream.com
Sun May 24 22:38:14 CDT 2015


Cory,

Are you using the TSO features of your nic on the host?  While I haven't 
heard or seen this exact problem, I would suspect that it's something to 
do Tcp Off-loading as the VM is likely sending 65k sized packets to the 
host machine and relying on the TSO features to break those large 
packets into frames for the wire. Given that you are seeing problems at 
65k I would really look hard at TSO.

You should be able to tell by running an ifconfig and looking for the 
TSO4 option.  Or at least that is what Google is telling me. My Freebsd 
server doesn't have TSO nic's so I can't verify it. ^^

             Josh


On 5/25/2015 2:04 AM, Cory Smelosky wrote:
> This is a bit odd...
>
> Can't for the life of me figure out what could be the cause.
>
> I have a VM host, running Proxmox to do KVM VMs, inside one of the KVM 
> VMs is some SIMH and KLH10 instances. Host NIC is e1000e
>
> FTP TO these systems - never from - stalls for quite awhile every 
> 65536 (134320 if doing 9000-byte MTU across the board - even on the 
> bridges) bytes.
>
> VMs are doing virtio-net, changing has no effect.  Changing OS also 
> has no effect. Nor does changing host OS or virt. software. Nor does 
> running KLH10 on the host itself.
>
> Switch doesn't show any buffer failures, only buffer misses. Client 
> doesn't show any packets dropped/lost.
>
> TCPDUMP shows periodic (although regular) checksum fails and has the 
> DF bit.
>
> This seems to ONLY affect ftp.  No firewalls are in the way, SCP to 
> the host is fine.
>
> VDE attaching to FreeBSD tap, KLH10 attaching to Linux TAP, KLH10 
> attaching to FreeBSD pcap...it doesn't matter.  FTP to the host is 
> fine, FTP to the VMs themselves is also fine.
>
> This issue HAS NOT always been present, and it does not appear for 
> others.
>
> Any ideas?  I'm completely stumped...and the only pointer was ... the 
> number 65536 which refers only to a TCP/UDP window size.
>
> Of note, I have an identical setup open TO THE PUBLIC INTERNET without 
> issue.


More information about the geeks mailing list