Re: Socket buffer sizes with autotuning

["Jerry Chu" <hkchu@google.com>]

Just for the record, in my netperf RR (1MB request/20B reply) over
1GE/tg3 test on top of 2.6.18, changing tcp_tso_win_divisor to 0
from the default of 3 gets the cwnd down to 695 from 1037. Applying
the bound check (left * tp->mss_cache < 65536) will get cwnd down
to 737, not 695. Still big improvement though. (Anyway this may be
moot after I fix the real culprit.)

Jerry

On Mon, Apr 28, 2008 at 4:28 PM, John Heffner <johnwheffner@gmail.com> wrote:
>
> On Mon, Apr 28, 2008 at 1:38 PM, Jerry Chu <hkchu@google.com> wrote:
> > > Try this series of patches, against net-next.
> >
> > Ah, you already know about this problem. Yes it does the trick.
>
> I had not actually known about this.  (Just cooked up the patches
> after I saw your mail.)  This looks to me like a clear bug.  After
> looking carefully at the code there, it seems there was another small
> problem, too, though not likely to have much effect.
>
> Dave, can you apply the patches I just sent?
>
> Thanks,
>  -John
>