NSFnet router performance (was Re: IRC Net Bandwidth)

[Chris Torek]

>In article <11378 at dog.ee.lbl.gov> I wrote:
>>At the time the NSS routers peaked out at ... under 1000 packets per second.

In article <706 at seqp4.UUCP> jdarcy at seqp4.ORG (Jeffrey d'Arcy) writes:
>Ugh!  I knew the routers were inefficient, but this level of nonperformance
>defies comprehension.

Indeed.  One hopes it has been improved since then.  (In fact, I think
I remember something that implies it has: the MCI boxes that split each
T1 into three virtual cables, and---for performance---often bypassed
the NSS routers anyway [and note that the MCI boxes had no trouble
keeping up with T1 links], have been removed.)

I was just looking at a previous Usenix conference proceedings (Winter
1990), in which a paper on NSFnet traffic characterization noted that a
1-millisecond time stamp was sufficient because the machines could not
receive more than one packet, regardless of length, in that time.  This
also implies `under 1000 packets per second', so perhaps my memory was
not far off....

In any case, there is no real excuse not to be able to route at
Ethernet speeds or better, these days.  Given the fact that > 1Gbps
optical links will be in place any day now, routing performance is
going to be critical.  I am on the `IRC side' (if there can be said to
be such a thing) in this matter; I just refuse to pretend that current
systems make IRC (or indeed any small-packet system) as cheap as it
should be.
-- 
In-Real-Life: Chris Torek, Lawrence Berkeley Lab CSE/EE (+1 415 486 5427)
Berkeley, CA		Domain:	torek at ee.lbl.gov