[c-nsp] mitigate output drops on Cisco 2960G platform

[Phil Mayers]

On 09/21/2013 01:40 AM, Blake Dunlap wrote:
> Making the linux box slower in the ways described (other than packet
> pacing) would just increase the cpu load. If those are feeder servers,
> you're much better off dropping the link speed to 100m or configuring TC
> outbound than turning off a lot of the cpu offloading.

If packet pacing is OK, then disabling TSO should be OK too surely, 
since they will result in similar results from a CPU load PoV - the CPU 
preparing the individual TCP segments, not the NIC.

FWIW, my experience suggests that TSO makes relatively little difference 
to CPU load except in certain edge cases, such as very old/slow CPU 
paired with modern/fast NIC, high-throughput LAN apps like iSCSI, or 
artificial traffic like iperf.

I assert that, regardless of your link speed, the microbursts generated 
by TSO are sub-optimal, and that marginally higher CPU use is a price 
worth paying for better network-layer behaviour.

Regarding "tc" I'm not sure how this interacts with TSO; I think it 
would still result in a single large frame being passed to NIC, and thus 
a microburst - but "tc" would space out the large frames.

> Or alternately, not trying to connect 10 servers at 1g to a budget access
> layer switch with a single 1g uplink. If you don't want oversubscription,
> don't oversubscribe....

On that we agree ;o)