Fastest Linux Router to date?

Does anybody know any sources of information or
ballpark figures to what is the routing speed
of the fastest Linux router that could be constructed today
using the hotest PC (even if its a Quad CPU or multi-PCI bus
Dell,etc) and using high end "server" NICs, etc (though
not using specialized router hardware).

How many orders of magnitude would this be slower than
an evil CISCRAP unprogrammable router, yada yada?

Thanks

-=-
This message was posted via two or more anonymous remailing services.

["Followup-To:" header set to comp.os.linux.]
LinuxRouterQuery enlightened us with:
> Does anybody know any sources of information or
> ballpark figures to what is the routing speed
> of the fastest Linux router

Fast in what way? Total bandwidth? Latency? Something else?

> How many orders of magnitude would this be slower than
> an evil CISCRAP unprogrammable router, yada yada?

Why do you think it'll be slower?

> This message was posted via two or more anonymous remailing services.

My my, aren't we paranoia?

Sybren
--
The problem with the world is stupidity. Not saying there should be a
capital punishment for stupidity, but why don't we just take the
safety labels off of everything and let the problem solve itself?

LinuxRouterQuery wrote:
> Does anybody know any sources of information or
> ballpark figures to what is the routing speed
> of the fastest Linux router that could be constructed today
> using the hotest PC (even if its a Quad CPU or multi-PCI bus
> Dell,etc) and using high end "server" NICs, etc (though
> not using specialized router hardware).

Routing is not very CPU intensive. When testing an older version of the Freesco
router software I was able to saturate the two 10 Mbps NICs with a 33 Mhz '486
and 8 MB RAM! So, it should not take much of a Pentium or Pentium 2 systems to
saturate 100 Mbps NICs.

> How many orders of magnitude would this be slower than
> an evil CISCRAP unprogrammable router, yada yada?

0

> Thanks
>
> -=-
> This message was posted via two or more anonymous remailing services.

No guts, no glory....


--
Phil Frisbie, Jr.
Hawk Software
http://www.hawksoft.com

Phil Frisbie, Jr. wrote:

> LinuxRouterQuery wrote:
>> Does anybody know any sources of information or
>> ballpark figures to what is the routing speed
>> of the fastest Linux router that could be constructed today
>> using the hotest PC (even if its a Quad CPU or multi-PCI bus
>> Dell,etc) and using high end "server" NICs, etc (though
>> not using specialized router hardware).
>
> Routing is not very CPU intensive. When testing an older version of the
> Freesco router software I was able to saturate the two 10 Mbps NICs with a
> 33 Mhz '486 and 8 MB RAM! So, it should not take much of a Pentium or
> Pentium 2 systems to saturate 100 Mbps NICs.
how did you measure that the links were saturated? iptables?
i'd like to see the rate at which data flows into or out of a nic in real
time
>
>> How many orders of magnitude would this be slower than
>> an evil CISCRAP unprogrammable router, yada yada?
>
> 0
>
>> Thanks
>>
>> -=-
>> This message was posted via two or more anonymous remailing services.
>
> No guts, no glory....
>
>

On Thu, 05 Feb 2004 21:12:07 +0000, Phil Frisbie, Jr. wrote:

> LinuxRouterQuery wrote:
>> Does anybody know any sources of information or ballpark figures to what
>> is the routing speed of the fastest Linux router that could be
>> constructed today using the hotest PC (even if its a Quad CPU or
>> multi-PCI bus Dell,etc) and using high end "server" NICs, etc (though
>> not using specialized router hardware).
>
> Routing is not very CPU intensive. When testing an older version of the
> Freesco router software I was able to saturate the two 10 Mbps NICs with a
> 33 Mhz '486 and 8 MB RAM! So, it should not take much of a Pentium or
> Pentium 2 systems to saturate 100 Mbps NICs.
>
>> How many orders of magnitude would this be slower than an evil CISCRAP
>> unprogrammable router, yada yada?
>
> 0
>
>> Thanks
>>
>> -=-
>> This message was posted via two or more anonymous remailing services.
>
> No guts, no glory....

I think i can assure you you're very wrong, what's 10Mbps to a router on a
really loaded network ? Or when being used to route a ddos the size to
MyDoom ? Routing itself might indeed not be very cpu intensive but when
you're talking the size of router the question itself suggests we're
talking a much higher scale of possible load.

For example, it is very possible to simply melt a network-card in a router
by applying access-lists to the routes, for example to counter a ddos.

At it's best you'd need a smp-box where even the extra instructions like
mmx, sse, 3dnow/pro are used to improve routing and/or access-lists to be
able to talk about a top of the line router performance.

Also, you can't say to measure something in a two-point network-system,
you'de be talking more then two subnets, if possible quite a bit more to
get some real performance indication.

Maybe you should verify the kind of cpu cisco uses in their routers ...
Some of the backplanes-cpu's are the highest-clocked cpu's available ...

--
Best Regards,

Dr. Chandra

"Dr. Chandra" <drchandra@_LOSE-SP4M_pandora.be> wrote in message news:<pan.2004.02.06.14.41.49.270252@_LOSE-SP4M_pandora.be>...
> On Thu, 05 Feb 2004 21:12:07 +0000, Phil Frisbie, Jr. wrote:
>
> > LinuxRouterQuery wrote:
> >> Does anybody know any sources of information or ballpark figures to what
> >> is the routing speed of the fastest Linux router that could be
> >> constructed today using the hotest PC (even if its a Quad CPU or
> >> multi-PCI bus Dell,etc) and using high end "server" NICs, etc (though
> >> not using specialized router hardware).
> >
> > Routing is not very CPU intensive. When testing an older version of the
> > Freesco router software I was able to saturate the two 10 Mbps NICs with a
> > 33 Mhz '486 and 8 MB RAM! So, it should not take much of a Pentium or
> > Pentium 2 systems to saturate 100 Mbps NICs.

You've not worked with very large tables or ACLs then, not to mention
a box stuffed to the gills with multi-port cards serving a number of
subnets. And then there are voice and video streaming, buffering of
data streams, cacheing routing info in separate tables for each
protocol enabled, and queueing manipulation to achieve QoS and RSVP
requests. The list continues ...

> >
> >> How many orders of magnitude would this be slower than an evil CISCRAP
> >> unprogrammable router, yada yada?

[snip]

> I think i can assure you you're very wrong, what's 10Mbps to a router on a
> really loaded network ? Or when being used to route a ddos the size to
> MyDoom ? Routing itself might indeed not be very cpu intensive but when
> you're talking the size of router the question itself suggests we're
> talking a much higher scale of possible load.
>
> For example, it is very possible to simply melt a network-card in a router
> by applying access-lists to the routes, for example to counter a ddos.
>
> At it's best you'd need a smp-box where even the extra instructions like
> mmx, sse, 3dnow/pro are used to improve routing and/or access-lists to be
> able to talk about a top of the line router performance.

This makes no sense to me whatsoever. These are manufacturer/chip
architecture specific re: manipulating/extending floating point
operations used in graphics _rendering_ instructions used by advanced
rendering libraries. Routers don't even look at the data payload and
certainly don't do any rendering. The extended _registers_ can be
used for dog-slow integer manipulations but a router doesn't need/use
them (and not much else that I'm aware of).

> Also, you can't say to measure something in a two-point network-system,
> you'de be talking more then two subnets, if possible quite a bit more to
> get some real performance indication.
>
> Maybe you should verify the kind of cpu cisco uses in their routers ...
> Some of the backplanes-cpu's are the highest-clocked cpu's available ...

Except as noted above, your points are well taken, but I think OP is
interested in latency and how closely the router can achieve near wire
speed throughput. With a general purpose OS/router the # of NICs will
certainly be a factor as all those interupts must be processed. Also
note that x86 memory (ie., ram) access is about the slowest devised by
man -- thus the need for large on-chip caches and ~3ghz clocking to
achieve what others achieve with 1/4 to 1/6 the clock speed.

Dedicated routers will always (when properly configured) win on this
score as they use ASICs at _each_ port to boost throughput and reduce
latency. They also use a much more advanced technology for internal
switching/forwarding between ports. Doubt the IOS is substantioally
"faster" than a properly configured/stripped Linux kernel used just
for routing.

Then there is the question of Cisco's proprietary routing protocols,
especially EIGRP, and the possible need to run several routing
protocols over several network architectures. Network performance is
all that matters; not any one measure of "speed". Eg., how fast does
the net converge after a change? Is redundacy built into the network
and if so how are the "redundant" paths utilized to distribute/balance
the load? How much work/time do you have to administer your network?
How much money to build it? Most important of all, how fast is fast
enough considering your constraints. Taken together, Linux based
routing wins out over the proprietary offerings in an increasing # of
environments, regardless of any "speed" measure.

Computing, especially network computing, is not a 100 meter dash; it's
a marathon where some jerk keeps moving the finish line farther away
every time it comes into view.

m2c
prg
email above disabled

P Gentry wrote:

> "Dr. Chandra" <drchandra@_LOSE-SP4M_pandora.be> wrote in message news:<pan.2004.02.06.14.41.49.270252@_LOSE-SP4M_pandora.be>...
>
>>On Thu, 05 Feb 2004 21:12:07 +0000, Phil Frisbie, Jr. wrote:
>>
>>
>>>LinuxRouterQuery wrote:
>>>
>>>>Does anybody know any sources of information or ballpark figures to what
>>>>is the routing speed of the fastest Linux router that could be
>>>>constructed today using the hotest PC (even if its a Quad CPU or
>>>>multi-PCI bus Dell,etc) and using high end "server" NICs, etc (though
>>>>not using specialized router hardware).
>>>
>>>Routing is not very CPU intensive. When testing an older version of the
>>>Freesco router software I was able to saturate the two 10 Mbps NICs with a
>>>33 Mhz '486 and 8 MB RAM! So, it should not take much of a Pentium or
>>>Pentium 2 systems to saturate 100 Mbps NICs.
>
>
> You've not worked with very large tables or ACLs then, not to mention
> a box stuffed to the gills with multi-port cards serving a number of
> subnets. And then there are voice and video streaming, buffering of
> data streams, cacheing routing info in separate tables for each
> protocol enabled, and queueing manipulation to achieve QoS and RSVP
> requests. The list continues ...

That is a very good point. The router was being used with just three NICs for a
small business; two local LANs and the WAN connection. There was port forwarding
from the WAN to one of the LANs for a couple of servers, and the other LAN had
about 10 clients with a short list of blocked outgoing ports.

--
Phil Frisbie, Jr.
Hawk Software
http://www.hawksoft.com