network virtualization performances

[Daniel Lezcano <dlezcano@fr.ibm.com>]

Hi all,

I am currently working on the network virtualization for lightweight 
containers based on the network namespace concept.

The network device list are relative to a namespace, so in one 
namespace, you can not access to network device belonging to another 
namespace. To allow the namespace to communicate, a special network 
device has been written. It is like a "pipe" but for the network with an 
extremity in each namespace. By this way, the virtualized environment 
can send traffic via a network device (an extremity of the "pipe") and 
the traffic goes to the namespace where the real nic is.

During some performance testing, I noticed that using the Gigabyte NIC 
and netperf, the cpu usage is much more important when the network is 
virtualized than when not. After investigating, it appears the real NIC 
is an offload card and when we are using the virtual network device, the 
offloading is lost and the kernel spends a lot of time to compute 
checksums and tcp segmentations. That makes sense because the virtual 
device has not the HW feature like the real NIC.

After looking in internet, I found the "Optimizing Network 
Virtualization in Xen" document and the problems seems to be exactly the 
same.

Do you have some pointers, informations about these optimization, 
specific optimization code or anything else which can help me ?

Thanks you in advance.

Regards.

   -- Daniel