Re: [SAGE] Enterprise Documentation

[Gene Rackow <rackow@mcs.anl.gov>]

Lots of this comes from experience and testing.  A number of years
ago I was working on designing an I/O server system and determining
what sort of controllers were needed to match the required specs
was difficult.  It's been some time so I can't give the specifics,
but we had 2 network cards to test by one manufacturer.  The cards
were but pre-releases.  We wanted to put 4 to 6 cards in server.

Both cards could saturate the network.  Now is where the fun begins.
The first card costs about $200, and was fairly stupid by itself,
and therefore took about 60% of the CPU to push those packets.
You could realisticly only put 2 of these cards into the system before
you overloaded the CPU and additional interfaces reduce the performance
of all cards installed.  The server had plenty of backplane bandwidth
to handle the load it was all in the CPU having to handle all the packet
formation, checksums, etc.

The 2nd card costs about $1500 and took less than 5% of the CPU to
saturate the network.  This card showed real promise for what we
needed to do.  With this card, there were not enough slots in the
backplane to saturate the CPU on I/O alone and would have been a
great thing to put into the machine.  The problem is the manufacturer
decided that it was not worth taking that card into production since
the expected price would limit the sales to the point they didn't think
they would recover the design costs.

We see some of the same sorts of things happening with cards today.
It may very well be the netcard that you are using is on the lower
end of performance as well.

The company really gave you a very simplified answer, but there is probably
much more going on than just the faster net card.  With the faster net, you
probably changed the load, you are probably doing more NFS processing
as well.  This would cause more of a load on the CPU.  This may also 
change the way the disks are being accessed and therefore more waiting
for disks to spin, etc.  

--Gene

 >
 >> Then when I didn't know where else to turn I called the company of the
 >product and there 
 >> response was that since you switch over to gigabit your network bandwidth
 >has increased 10 folds
 >> there for the faster the traffic the more instense the proccesor has to
 >grind. 
 >> ...
 >>  how does one begin to understand such topics?
 >