Re: Much improved sendfile(2) kernel implementation

David Malone wrote:
> On Fri, Sep 22, 2006 at 11:48:23PM +0100, Robert Watson wrote:
>> The impact of TSO is clearly dramatic, especially when combined with the 
>> patch, but I'm a bit concerned by the drop in performance in the patched 
>> non-TSO case.  For network cards which will always have TSO enabled, this 
>> isn't an issue, but do we see a similar affect for drivers without TSO?  
>> What can we put this drop down to?
> 
> We probably also need to make sure that any performance increase
> in TSO isn't due to us getting TCP congestion control wrong. I think
> in Linux they had problems when they first introduced TSO because
> TCP was advancing the congestion window by a TSO-sized chunk instead
> of a wire packet. OTOH, I think Andre and Drew's tests are low-latency,
> so congestion control isn't likely to be playing a big role, so the
> improvements are unlikely to be due to this.

The congestion window is increased based on the ACK's received.  TSO
is only done on the send side and only up to the current congestion
window.  I have been careful not to get any changes in congestion
control behavior with TSO.  (Which does not mean that there may be
other bugs lurking in our congestion control.)

-- 
Andre