[Bloat] debloat-testing tree rebased, some patches dropped

[Jonathan Morton]

On 30 Mar, 2011, at 11:06 pm, Dave Taht wrote:

> TCP-FIT, anyone?

Well, I know enough about it to be able to implement it, but are we allowed to?

I also measured some asymmetry between two of my machines, which led me to the Sun GEM driver.  The GEM is technically a GigE NIC, but in the machine I'm testing (a PowerBook G3) it has a 100base-TX PHY attached.  Regardless, the GEM driver in Linux sets the ring buffer to 128 entries for both RX and TX.  There is an obvious setting for 32 entries for TX.

The other machine has a VIA Rhine NIC, and the driver for that implies that the ring buffer has 16 entries but deliberately limits it to having only 10 entries in use.  The Responsiveness figures are much higher when all the traffic comes from the Rhine, compared to when any traffic comes from the GEM.

 - Jonathan