[rescue] Total corporate madness (

Francisco Javier Mesa-Martinez lefa at ucsc.edu
Fri Aug 8 00:00:19 CDT 2003

On Thu, 7 Aug 2003, Lionel Peterson wrote:
> Well, speaking from ignorance, and grossly simplifying (cause that the
> level of understanding I have on these topics), it seems to me that
> since:
>  a) A CPU can only do one thing at a time
>  b) Only one packet can be on a network segment at one time

Yeah, the problem is that your ethernet NIC presents requests for service
at a rate that is a few orders of magnitude less than the actual service
rate for the main server in your system, in this case the CPU. Besides
most of those requests are DMA anyways.

Servers are not CPU buound by any means, unless you are doing some sort of
exoteric backend stuff, a modern single CPU system with well valanced IO
will not be saturated by a few NICs by a loooooong shot :).

> It follows that if I had two nics on a single CPU system, and if there
> were file requests coming in on each nic, one would wait for the
> previous one to complete.

Your bottleneck is the PCI NIC and the harddrive if you are dealing with
large files, or your memory if you are dealing with on memory files.
Hardly taxing on the CPU, since most of the requests will be serviced
through DMA.

> If each nic had it's own CPU, then requests that come in can be
> dispatched handily. Oh, and to follow my simplification, you would want
> lots of drives on lots of controllers, and each drive would have unique
> files that would seldom/never be needed at the same time (spread the
> load across drive spindles/controllers)...
> Remember, here on this list, excess is assumed ;^)


More information about the rescue mailing list