Re: IPS Issues document

[julian_satran@il.ibm.com]

I would like to clarify my comment. I am not contesting the fact that CPU
will get saturated.
My concern is that CPU load decrease does not justify specialized silicon
for protocol processing
(as it gets soon out-dated like the 186 NICS) and you have the nightmare of
supporting 2 protocol stacks.  Memory load on the other hand can't be
solved any other way than by off-loading and reducing copies to a minimum.

Regards,
Julo

Zachary Amsden <zamsden@cthulhu.engr.sgi.com> on 14/03/2000 20:53:41

Please respond to Zachary Amsden <zamsden@cthulhu.engr.sgi.com>

To:   "Bradley, Mark" <mark_bradley@btc.adaptec.com>
cc:   ips@ece.cmu.edu (bcc: Julian Satran/Haifa/IBM)
Subject:  Re: IPS Issues document




> Julo, good comments.  My observation is that faster CPU's
> don't solve the problem.  We have been getting faster CPU's
> since the beginning of time and the only semi-constant is
> that we find that the CPU continues to get utilized.  Large
> MP machines are bought because the processing horsepower
> is needed for some appl.(s) that consumes it.  Therefore they
> don't have 'extra' cycles to spare.  Most generic OS's
> don't support dedicating a processor to specific jobs,
> either.  Putting a small processor someplace (an IOP) else
> typically doesn't scale as system processing speed increases,
> so those often become bottle-necks after some time.  Remember
> those 186-based networking cards?
>
> Bottom line is that % CPU utilization has been, is, and will
> continue to be an important consideration for some time.
> It seems we never have enough CPU, memory or disk.  We seem
> always to find ways to use it up.  (I remember years ago
> thinking "Wow!  A megabyte on a single platter!  We'll have
> enough disk space to last 3 or 4 years!!!"  I also recall
> thinking that a 68020 made for a pretty fast workstation.
> Now I complain that my dual 450 MHz PC with 128 MB of memory
> is not fast enough to be a decent office machine next year...
> :{)   )

I agree completely with that - protocol processing and interrupt time alone
for a 1 Gbps connection can easily saturate a single CPU, even with
interrupt
coalescing and many other tricks.  When you move to a 10 Gbps connection,
you
can't just add more CPUs to deal with the problem, because of both cost and
the fact that you don't get a perfect speedup as you add more CPUs.  This
makes techniques that reduce CPU and bus utilization more valuable as
network
speeds increase.

--
Zachary Amsden  zamsden@engr.sgi.com  3-6919  31-2-510  Core Protocols