Subject: Re: Slow UDP traffic
To: None <netbsd-users@NetBSD.org>
From: Marc Tooley <netbsdMLpostNO@SPAM.quake.ca>
Date: 06/11/2004 09:35:17
On Thursday 10 June 2004 14:51, Steven M. Bellovin wrote:
> I don't care how clever your tricks are; there's just not a lot of
> room for improvement....
If you're talking about a local link, then the fastest way to transmit
meaningless data is to eliminate as much overhead as possible and
simply blast it through. That's what he's (you're?) testing isn't it?
It's very possible to do better than TCP with meaningful data, and is
being done right now by companies like Data Expedition--there are real,
measurable gains over standard TCP because they've relaxed the
constraints that TCP operates under. I suppose if you don't believe me
that's your decision.
You're right: it's difficult to do better than TCP *under similar
constraints*--it's about as optimized as it's going to get. But once
you relax those constraints and shift some others onto applications,
interesting speeds are achieveable.
> But the moral is simple: measure
> before you optimize; you may find that the bottleneck is somewhere
> else entirely.
---exactly! Like in socket options, or buffer sizes. Sometimes fiddling
with those a little can yield very acceptable returns.