Subject: Re: Multiple net interfaces
To: Greg S. Burd <>
From: Dirk Steinberg <>
List: current-users
Date: 03/02/1994 09:23:50
>>>>> "Greg" == Greg S Burd <> writes:

    Greg> Robert Shady Writes: Right.  So does the Livingston
    Greg> Portmaster Terminal Server, but I don't have one of those,
    Greg> hence the question.  Somehow, there must be a way to connect
    Greg> two SLIP/PPP devices to a dial-up SLIP/PPP provider and have
    Greg> 2*(modem speed) throughput (or more!).

    Greg> This is impossible for several reasons.  First, the nature
    Greg> of IP routing.

This is not true. You certainly can have alternate routes. It's just a
questions of whether you software supports this. You need some kind of
dynamic load balancing.

    Greg> Second, the speed at which your system can operate.
    Greg> SLIP and PPP generate a huge amount of overhead
    Greg> even with a really good serial card.  Unless you get a
    Greg> serial card which does DMA, VESA, or PCI data transfers, the
    Greg> processing overhead will saturate your 486-DX4 or even your
    Greg> Pentium.

Nonsense. One 38400 BPS link will certainly *not* saturate a 486DX2/66
or a Pentium, especially if your UART has a built-in FIFO (16550), at
least not under Linux. I have used 38400 BPS *without* FIFOs (16450)
on a 386DX/40 Linux machine without loosing characters, and the CPU
was *not* saturated in any way. The only problem can be interrupt
latency, but, as said above, this can be eased a lot by using FIFOs.
Anyway this is no argument to prove that it cannot work, it just says
that you have inferior hardware and poorly written software. You
certainly don't need 32 bit busmaster DMA to handle more than one
38400 BPS link (which is just 4.7 *kilobytes* per second). Any stock
ISA bus has a bandwidth that is at least 1000 times as high. (This is
not to say that you could have 1000 links on one ISA bus, which you
certainly cannot.)

    Greg> 38400 BPS is lot of information to move across an
    Greg> 8, 16, or even 32 bit bus with all the interrupts, etc.
    Greg> Even if the serial driver is "smart" enough to do larger
    Greg> transfers, the overhead is a problem.  This is why
    Greg> MorningStar Technologies has something called SnapLink(tm).
    Greg> It is a SCSI serial device designed to handle T1 bandwidth.

Don't know about MorningStar stuff, so I cannot comment.


Dirk W. Steinberg - RWTH Aachen - Internet email:
Aachen University of Technology / IS2-Integrated Systems in Signal Processing
Rhein.Westf.Tech.Hochsch. Aachen / Integrierte Systeme der Signalverarbeitung
Templergraben 55 / D-52056 Aachen / phone:+49 241 807879 / fax:+49 241 807631
Home address: Kleikstr. 63, D-52134 Herzogenrath,Germany/phone: +49 2406 7225