[9fans] having the Nagle cake and eating it too

[9fans] having the Nagle cake and eating it too

[Mike Haertel]

The problem with Nagle seems to be that it clobbers performance
of RPC applications.  The communications pattern of such an
application is almost always:

	write(...request...);
	read(...response...);

(for clients) or

	write(...response...)
	read(...request...)

for servers.

Why not have read() force any outstanding writes to the network?
That way you could do the right thing even with programs whose
authors are too naive to know about turning off Nagle.

Re: [9fans] having the Nagle cake and eating it too

[Thomas Bushnell, BSG]

mike@ducky.net (Mike Haertel) writes:

> Why not have read() force any outstanding writes to the network?
> That way you could do the right thing even with programs whose
> authors are too naive to know about turning off Nagle.

That works fine for RPC programs, but it kill the Nagle algorithm for
programs that have continuous reads going on.

Re: [9fans] having the Nagle cake and eating it too

[presotto]

[-- Attachment #1: Type: text/plain, Size: 541 bytes --]

Tried that.  Things like telnet constantly read.  It works
sometimes, not others, ...

I tried various ways to ``guess'' what the app was doing and
either apply Nagle or not.  I could get it to work pretty
regularly with 9P but it would break with other things.

I finally gave up and took it out since at least once a month
it would brak something.  I tried it as an option for a while
but the only thing anyone wanted it for was telnet and we just
don't do that anymore, ssh has replaced it and our window system
line buffers.

[-- Attachment #2: Type: message/rfc822, Size: 1702 bytes --]

From: Mike Haertel <mike@ducky.net>
To: 9fans@cse.psu.edu
Subject: [9fans] having the Nagle cake and eating it too
Date: Mon, 26 Nov 2001 11:29:08 -0800 (PST)
Message-ID: <200111261929.fAQJT8f01844@ducky.net>

The problem with Nagle seems to be that it clobbers performance
of RPC applications.  The communications pattern of such an
application is almost always:

	write(...request...);
	read(...response...);

(for clients) or

	write(...response...)
	read(...request...)

for servers.

Why not have read() force any outstanding writes to the network?
That way you could do the right thing even with programs whose
authors are too naive to know about turning off Nagle.