Skip site navigation (1)Skip section navigation (2)
Date:      Thu, 28 Oct 2004 15:38:08 +0400
From:      Gleb Smirnoff <glebius@freebsd.org>
To:        Pawel Malachowski <pawmal-posting@freebsd.lublin.pl>
Cc:        freebsd-net@freebsd.org
Subject:   Re: PPTP/PPPoE mpd/poptop performance
Message-ID:  <20041028113808.GB50262@cell.sick.ru>
In-Reply-To: <20041026153108.GA91134@shellma.zin.lublin.pl>
References:  <20041026153108.GA91134@shellma.zin.lublin.pl>

next in thread | previous in thread | raw e-mail | index | archive | help
On Tue, Oct 26, 2004 at 05:31:08PM +0200, Pawel Malachowski wrote:
P> I would like to ask people using mpd about performance on particular hardware
P> setups. I am interested in the numbers of sessions (probably PPTP with weak
P> encryption) and total bandwith, that can be achieved with, e.g.:
P> . 300MHz CPU,
P> . 1GHz CPU,
P> . 2GHz CPU.
P> Won't PPPoE behave better than PPTP? (My goals are authentication and
P> performace.)
P> I guess poptop will be much slower and is not a good solution for bigger
P> setup (since it works entirely in userland)?
P> I'm talking about small ISP environment, about 100-200 simultaneous
P> connections over LAN, about 5-10Mbit/s total bandwith.

If your going to use encryption, then it will be the bottleneck. And
even difference between kernel(mpd) and userland(poptop,pppoed)
implementations is not going to be significant. You will need as
much fast CPU as possible.

I'd suggest to choose PPPoE, not PPTP, because the latter is quite
complicated and violated by some client implementation. You will
not find any problems with PPPoE, since ng_pppoe is compatible with
all known PPPoE implementations.

I'm running two PPPoE servers with no encryption:

1) userland ppp, ~200 sessions, ~2Mbit/s. Under peak traffic load
   is significant.
2) mpd, 10 - 50 clients (a small net). The load is almost zero.
   Can handle wirespeed 100Mbit, with interrupt load equal to
   load on pure Ethernet routing.

I have no experience with number of clients > 200. I suppose we are
going to notice performance degradation in linear search functions
in netgraph or interface code only when number of clients >> 1000.
If someone has such requirements, I'd be happy to work with him.

-- 
Totus tuus, Glebius.
GLEBIUS-RIPN GLEB-RIPE



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20041028113808.GB50262>