Hi Gerry,
thanks for your input about the eepro. It sounds and interesting (and hot)
topic of debate .... if the problem is to be the 3com NIC, then the solution
should be as easy as testing the router with another eepro replacing the
3com.
Anyway... don't you think the performance hit is perhaps *too* big in order
to be a NIC issue ? I mean, I would see OK that eepro made ( let's say ) a
40mbit/sec routing, and the 3com stood at, for example, at 30. But standing
at 6'4 Mbps ?? Isn't that too low ? I personally find it hard to believe
that simply plugging a Eepro instead the 3com will boost the router to, say,
20 or 30 mbit/sec ...
OTHO, I've been told to change the HZ define in include/param.h from 100 (
default x86 ) to 1024 , in order to get faster task switching. Since we are
talking about transmitting LOTS of small packets, increasing kernel
granularity as my partner suggests could perhaps solve the problem. I'm
still setting up a backup router in order to test with this parameter... do
you think this measure could fix the problem ? Your opinion is welcome...
TIA!
Jose Miguel Varet
Tech. Dept. Chief
ATT, S.L.
----- Original Message -----
From: "Gerry Creager" <gerry@xxxxxxxxxxx>
To: "Jose Miguel Varet" <varet@xxxxxxxxx>
Sent: Tuesday, July 03, 2001 5:07 PM
Subject: Re: [LARTC] Routing performance
> Jose,
>
> I'd be looking at the NICs. I believe you will get better performance
> out of the Intel EEPro 100 cards than the 3Coms. This has been a
> subject of great debate on the Beowulf list (www.beowulf.org), but it
> was some time back.
>
> I'm using some Dell boxes with 2 EEPro100's and a DLink 4-port
> Tulip-based chipset. I've not metered very tightly, but we're seeing
> much better than 6-7 Mb/sec... We're using those as routers in a
> testbed environment simulating the Internet and have had a lot of
> traffic going thru them.
>
> Good luck, gerry
> --
> Jose Miguel Varet wrote:
> >
> > Hello all,
> >
> > I am using a linuxbox as one of my system's router ( I work on an
ISP ),
> > with the following specs :
> >
> > - PIII 1 Ghz
> > - 2.4.5 cbq-enabled kernel
> > - eth0 3com 905B , eth1 Intel Etherpro 100
> > - 256 MB Ram
> >
> > The traffic usually peaks at 6 Mbit/sec . Until here, it's all OK. I am
very
> > satisfied with this router, and I can place cbq's, filters, etc. etc.
> >
> > Recently we have been adding some new customers behind that router, and
the
> > traffic began to grow. To my surprise, when the router reachs certain
output
> > limit ( around 810 kBytes/sec or 6'5 Mbit/sec ) the output remains at
that
> > level *except* when it begins to decrease again. Looks like to me as if
the
> > router box couldn't handle _more_ than 6'5 mbit/sec , uh... I know that
> > cannot be so, since the system use doesn't go beyond 9% as show by the
"top"
> > command, and both NICs are good ones... I cannot believe the bottleneck
is
> > to be in the routing code.
> > But, in order to clarify the situation, I'd need your opinions about
this
> > issue, which would help me greatly. My real bets for the problem are :
> >
> > a) our carrier is limiting the upstream bandwith to our router, so no
more
> > external users can reach our webs provided a certain limit has been
reached
> > ( at 810 kBytes/sec , incoming traffic goes as high as 112 kBytes/sec .
> > Perhaps they are limiting us the incoming traffic.
> >
> > b) our web servers cannot output more bytes that those. For testing
this, I
> > temporaly took down one web server. Inmediately, the other went 100
> > kBytes/sec up. So they CAN have more output, individually. So my problem
is
> > to be on the router, or with my carrier and my upstream bandwidth.
> >
> > I am very annoyed with this, since I *do* believe that 2.4 kernel can
handle
> > MUCH more that a miserable 6'5 mbit/sec. Where could the problem be ?
Your
> > opinions are really welcome...
> > Many thanks in advance.
> >
> > Jose Miguel Varet
> > Tech. Dept. Chief
> > ATT, S.L.
> >
> > _______________________________________________
> > LARTC mailing list / LARTC@xxxxxxxxxxxxxxx
> > http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO:
http://ds9a.nl/2.4Routing/
>
> --
> Gerry Creager | Never ascribe to Malice that
> AATLT | which can adequately be
> Texas A&M University | explained by Stupidity.
> 979.458.4020 (Phone) | -- Lazarus Long
> 979.847.8578 (Fax)
>
