North American Network Operators Group
Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical

RE: Jumbo Frames (was Re: MAE-EAST Moving? from Tysons corner torestonVA. )

  • From: Roeland Meyer (E-mail)
  • Date: Mon Jun 19 18:56:42 2000 
> Marc Slemko: Monday, June 19, 2000 10:06 AM
>
> On Mon, 19 Jun 2000, Roeland M.J. Meyer wrote:
>
> > I should have re-caveated, for your benefit. I am not testing
> > with a bazillion-byte file. I am testing with query/response
> > against a RDBMS host. IOW, a typically real-world(tm)
practical
> > application. The responses range from 3-50KB, with anomalies
out
> > to 100KB. The slow-start algorithm has been identified as the
>
> Erm... no, then your problem is opening and closing TCP
> connections all the time.  Don't do that.

I don't have much choice there. Each query/response is a new
connection. Even SQLnet is limited with batch query optimization.

> It hurts you in a lot of other ways.

Yes, it does. I'm still scraping off the charred back-side meat.

> It really isn't appropriate to go around saying "you need
> larger MTUs to
> fill a 100 meg link, period" when you really mean "in one
particular
> situation where I am opening and closing TCP connections and
> only sending
> a very small amount of data over each, you need larger MTUs".

Hm, I don't remember the "period" and I thought that I'd outlined
my case a few messages back.

> I wouldn't be so quick to say slow start is useless, either.
> Perhaps with
> small window sizes, but as soon as they get big enough...

Here is where you may not have thought it through enough. On a
dedicated FDX link, what need is there for slow-start? Only the
transmitter and receiver are on the wire and the other-end has a
separate transmit circuit to talk back with (the other side of
the FDX link). If the switch can't keep up then I need a switch
that can. In this case, I happen to know that the switch is fine.
I'm feeding CAT5 straight from the switch to the NIC on the
server. The other side is similarly connected. Slow-start is a
legacy requirement for non-switched networks and is still exists
for legacy reasons. In switched FDX environments, it would be
real nice if I could just turn it off, as a configuration issue.
In fact, there's a lot of stuff that could probably be stripped
from a stack, for switched FDX environs and modern SMP hosts.
Even switched 100baseTX could benefit.