Mailing List Archive

BSD/386 can handle full routing table with some amount of tweaking.
Also, gated becomes essentially useless as soon as system starts
to do paging; so it's better 64Mb or more :-)

Reliability of 486 platforms... closer to non-existant. ciscos
do not cost $$$s for nothing. Also, ciscos are much more
manageable.

There are several real-life E-1 BSD/386 routers in the Internet,
though owners want to replace them with ciscos (capacity problems).

--vadim
- - - - - - - - - - - - - - - - -

From: bmanning@is.rice.edu (William Manning)
Subject: innocent remark

>From another list...

given that a BSD/386 or NetBSD system can be put together quite cheaply and
does a fine job at handling PPP and SLIP and multiple ethernets, this seems
like a fine alternative to dedicated routers like Cisco for T1/E1-speed
internet gateways. especially considering the power of GateD 3.5, and that
a 16MB cisco can't handle a "full internet route table" while a 16MB i486
machine can do it easily.
...

Comments from the gated crowd? The cisco crowd?

Go for it, have a party. I've got 2 BSD boxes myself, and I prefer them
over clunky cisco routers. Besides, have you ever been able to play
DOOM on an AGS?

--
Regards,
Bill Manning
- - - - - - - - - - - - - - - - -

> given that a BSD/386 or NetBSD system can be put together quite cheaply and
> does a fine job at handling PPP and SLIP and multiple ethernets, this seems
> like a fine alternative to dedicated routers like Cisco for T1/E1-speed
> internet gateways. especially considering the power of GateD 3.5, and that
> a 16MB cisco can't handle a "full internet route table" while a 16MB i486
> machine can do it easily.

The end of the last sentence is wrong (the start of the last sentence
might be too, though I'm biased enough to maybe believe it). I know
for a fact that, while gated's memory usage is not unreasonable, gated
code does tend to resolve memory/CPU tradeoffs in favour of spending
memory and saving CPU. And the per-route memory consumption in the
Net/II kernel forwarding table is way out of line with what a decent
router implementation should do. If you want your 486 box to compete
with a 16MB cisco you really need to buy it a couple of extra rows of
SIMMs. By the time you get enough neighbours and alternate paths in
there to overwhelm a 16MB Cisco the 16MB 486 box will be spending most
of its time trying to get large chunks of gated into and out of the
page space, though a 32MB 486 box should still be pretty chipper.

My (somewhat dated) observation is that i486 machines also seem to do
a whole lot of work to get quite mediocre packet-per-second forwarding
rates. On the other hand, if you buy it a good screen and interface card,
the graphics are much better than even a 64MB Cisco.

Dennis Ferguson
- - - - - - - - - - - - - - - - -

>> a 16MB cisco can't handle a "full internet route table" while a 16MB i486
>> machine can do it easily.

SL-IIJ-1>show ip b s
BGP table version is 9430997, main routing table version 9430997
19216 network entries (32713 paths) using 2945880 bytes of memory
917 BGP path attribute entries using 73740 bytes of memory
...

SL-IIJ-1>show mem free
Head FreeList Total(b) Used(b) Free(b) Largest(b)
Processor 310824 2E96D0 13315724 7699588 5616136 5131068
I/O 6000000 2F04B4 1047544 370708 676836 676836
SRAM 1000 2F11E8 65412 34760 30652 30652

SL-IIJ-1> show ver
...
cisco 4000 (68030) processor (revision 0xA0) with 16384K/1024K bytes of memory.
...

Cheers,
-drc
- - - - - - - - - - - - - - - - -

there are a number of companies that have made modest enhancements
(some of which you have mentioned) to NetBSD and are or will be deploying
routers based on the code using PC or other SBC "platforms".

I consider these people to be the "cloners" like a gateway or a dell.

They make wonderful 1e1T, 1e1B (BRI), and 1e1V (v.32bis)
"edge" routers, but like the PC cloners they will set their sights higher
and higher.

Marty
> > given that a BSD/386 or NetBSD system can be put together quite cheaply and
> > does a fine job at handling PPP and SLIP and multiple ethernets, this seems
> > like a fine alternative to dedicated routers like Cisco for T1/E1-speed
> > internet gateways. especially considering the power of GateD 3.5, and that
> > a 16MB cisco can't handle a "full internet route table" while a 16MB i486
> > machine can do it easily.
>
> The end of the last sentence is wrong (the start of the last sentence
> might be too, though I'm biased enough to maybe believe it). I know
> for a fact that, while gated's memory usage is not unreasonable, gated
> code does tend to resolve memory/CPU tradeoffs in favour of spending
> memory and saving CPU. And the per-route memory consumption in the
> Net/II kernel forwarding table is way out of line with what a decent
> router implementation should do. If you want your 486 box to compete
> with a 16MB cisco you really need to buy it a couple of extra rows of
> SIMMs. By the time you get enough neighbours and alternate paths in
> there to overwhelm a 16MB Cisco the 16MB 486 box will be spending most
> of its time trying to get large chunks of gated into and out of the
> page space, though a 32MB 486 box should still be pretty chipper.
>
> My (somewhat dated) observation is that i486 machines also seem to do
> a whole lot of work to get quite mediocre packet-per-second forwarding
> rates. On the other hand, if you buy it a good screen and interface card,
> the graphics are much better than even a 64MB Cisco.
>
> Dennis Ferguson
- - - - - - - - - - - - - - - - -

Bill, Vadim,

This discussion doesn't really belong on this list, but since Vadim
has told some untruths about gated needing more than 64MB of memory...

> BSD/386 can handle full routing table with some amount of tweaking.
> Also, gated becomes essentially useless as soon as system starts
> to do paging; so it's better 64Mb or more :-)

Most of the ANS routers run 32 MB and run gated. A few have 64 MB.
If we didn't have the scram utility holding it's own routing table
(used to load routes into smart cards), we could probably get away
with 32 MB almost everywhere. Yes - we do take full routing. The SGI
I am typing on has 32 MB and runs gated and takes full routing and
does full logging and I do compiles and other stuff with no problem.

> Reliability of 486 platforms... closer to non-existant. ciscos
> do not cost $$$s for nothing. Also, ciscos are much more
> manageable.

The reliability of a Cisco is better, but PCs are not as bad as you
say. The manageability issue, I disagree with. I'd take the BSDI box
any day on that count. Just try expressing the policy on one of our
major ENSS routers on a Cisco. A BSDI box doesn't need a host to tftp
boot images and configs from (on the odd chance that you ever update
your Cisco software or have a config that doesn't fit into NVRAM) and
doesn't need a host to tftp logs to (on the odd chance that anything
ever goes wrong in a network and you need logging or tracing).

> There are several real-life E-1 BSD/386 routers in the Internet,
> though owners want to replace them with ciscos (capacity problems).
>
> --vadim

I don't think a PC today can realistically handle even one T1 at full
capacity. Also PC ethernet cards are sufficiently miserable that you
can only put one in a PC and get it to use the full ethernet
bandwidth.

Makes a fine 56k router or a low end dial in PPP or SLIP server if you
can live with one ethernet. Beyond that, the PC hardware just isn't
there yet.

Cisco doesn't make a laptop. :-)

Curtis
- - - - - - - - - - - - - - - - -

Some real data, these are test results for throughput (max zero loss rate)
for the Novell multiprotocol router running on a 486 box. these results
are for ethernet to ethernet through a T1 line, the limiting factor is the
T1. There are more sults on hsdndev.harvard.edu in pub/ndtl.

Small packets run at about a half T1, large packets at about 75%. The
limit is not in the PC it is in the serial line interface. (the same PC
can run at more than 50Mb/sec fddi to fddi.)

Scott

0064 byte packets
total: 043891
duration: 000028
total rate: 001567 pps
packet 020: 043891 (001567 pps)
0128 byte packets
total: 029089
duration: 000028
total rate: 001038 pps
packet 020: 029089 (001038 pps)
0256 byte packets
total: 017951
duration: 000029
total rate: 000619 pps
packet 020: 017951 (000619 pps)
0512 byte packets
total: 009585
duration: 000028
total rate: 000342 pps
packet 020: 009585 (000342 pps)
0768 byte packets
total: 006483
duration: 000028
total rate: 000231 pps
packet 020: 006483 (000231 pps)
1024 byte packets
total: 004172
duration: 000028
total rate: 000149 pps
packet 020: 004172 (000149 pps)
1280 byte packets
total: 003509
duration: 000029
total rate: 000121 pps
packet 020: 003509 (000121 pps)
1518 byte packets
total: 002958
duration: 000029
total rate: 000102 pps
packet 020: 002958 (000102 pps)

- - - - - - - - - - - - - - - - -

19216 network entries (32713 paths) using 2945880 bytes of memory
Head FreeList Total(b) Used(b) Free(b)
Processor 310824 2E96D0 13315724 7699588 5616136

I think that it's time for folks on these mailing lists (and on the
BGP list) to stop discussing the number of routes. It's really a
bogus metric. Memory (at least in our implementation) is consumed
for multiple reasons: paths, the AS path table, the routing table,
etc.

I would suggest that a much better characterization of BGP memory
usage is in terms of both number of routes AND number of paths. A
router with 20k routes and 20k paths uses a lot less memory than 20k
routes and 100k paths. Admittedly this isn't perfect, but it's a lot
better than just one dimension.

We now return you to your normally scheduled confusion...

Tony
- - - - - - - - - - - - - - - - -

>This discussion doesn't really belong on this list, but since Vadim
>has told some untruths about gated needing more than 64MB of memory...

I told that 16Mb is not enough. Memory for PCs is cheap and
it's better to be safe than sorry.

>The reliability of a Cisco is better, but PCs are not as bad as you
>say.

If you're lucky. I've seen a lot of burned PCs but only few
ciscos with problems. An average PC from Pretzelz'N'Computerz
simply won't work with Unix without massive tweaking.

>The manageability issue, I disagree with. I'd take the BSDI box
>any day on that count.

Yeah. Sorry, how do i do "show ip bgp summary" on BSD/386?

>Just try expressing the policy on one of our
>major ENSS routers on a Cisco.

That's the problem of the policy (though it's entirely different
kettle of fish). Sure, cisco can't handle real big access lists.

>A BSDI box doesn't need a host to tftp
>boot images and configs from

What is the average failure rate of cheap PC HD drives as
compared to flash ROM? What do you do if after power failure
fsck clears /etc? Sure, you can find an industrial-design PC
but it'll cost not less than a cisco.

>(on the odd chance that you ever update
>your Cisco software or have a config that doesn't fit into NVRAM) and
>doesn't need a host to tftp logs to (on the odd chance that anything
>ever goes wrong in a network and you need logging or tracing).

Never saw a configuration which does not fit into NVRAM. KISS.

> There are several real-life E-1 BSD/386 routers in the Internet,
> though owners want to replace them with ciscos (capacity problems).

>I don't think a PC today can realistically handle even one T1 at full
>capacity.

I only wanted to point our that there *are* such BSDI-based routers.
If you do traceroute to sovcom.kiae.su you'll see three of them.

>Also PC ethernet cards are sufficiently miserable that you
>can only put one in a PC and get it to use the full ethernet
>bandwidth.

Take a look at 3c509. It starts giving you data while the rest
of a packet is still in the wire. Or, better yet, EISA model.

>Makes a fine 56k router or a low end dial in PPP or SLIP server if you
>can live with one ethernet. Beyond that, the PC hardware just isn't
>there yet.

Hey, i wrote most Ethernet drivers, all sync serial drivers and PPP stuff
when i worked for BSD Inc. IMHO, you *can* do 2-3 port T-1 on a PC.

>Cisco doesn't make a laptop. :-)

Yeah, it also doesn't have joystick and soundblaster :-) :-) :-)

--vadim
- - - - - - - - - - - - - - - - -

> Yeah. Sorry, how do i do "show ip bgp summary" on BSD/386?

I agree this is a big defect. The traditional way of getting information
from gated has long been outgrown and SNMP mostly sucks for anything other
than turning something on a network monitor red. Being able to type
things and see specific interesting and entirely implementation-dependent
bits of internal state in response is neat, so this defect is being fixed.

Dennis Ferguson
- - - - - - - - - - - - - - - - -

From: bmanning@is.rice.edu (William Manning)
Date: Wed, 27 Apr 94 17:09:17 CDT
X-Mailer: ELM [version 2.3 PL11]

>From another list...

given that a BSD/386 or NetBSD system can be put together quite cheaply and
does a fine job at handling PPP and SLIP and multiple ethernets, this seems
like a fine alternative to dedicated routers like Cisco for T1/E1-speed
internet gateways. especially considering the power of GateD 3.5, and that
a 16MB cisco can't handle a "full internet route table" while a 16MB i486
machine can do it easily.
...

Comments from the gated crowd? The cisco crowd?

I would have to wonder about the forwarding performance of 16MB i486
BSD machine acting as a router. Also I am not sure what a full
internet route table should be, but a well-architected 16MB turnkey
router should be able to support a route table of well over 100,000
routes.

--
Regards,
Bill Manning

Joachim Martillo
Manager of Internetworking Research
Penril Datability Networks
Penril Datability Advanced Communications Research Center
190 N. Main St.
Natick, MA 01760
VOICE 508-653-5313
FAX 508-653-6415
EMAIL martillo@dss.com
martillo@penril.com

- - - - - - - - - - - - - - - - -