Re: Scyld vs kernel drivers

On Thu, 5 Dec 2002, Bogdan Costescu wrote:

> > I was wondering about the differences between Donald Becker's driver and
> > the driver that comes with the linux kernel,
...
> > especially the differences performance-wise.
> 
> I don't think that the driver (either of them) limits performance in any 
> way when using a modern card (905B or later) with a pretty modern CPU (I 
> normally test here with a 450 MHz CPU). I wasn't able to see any 
> difference in transfer rate between CPUs with 400 and 1200 MHz for my 

There is a difference in the interrupt pattern.  This can have a
throughput vs. latency performance impact on specific workloads, but
it's a very minor tradeoff, not a better/worse situation.
Now if I worked for Gartner or Mindcraft were paid, I could cook up a
benchmark and do an expanded-scale inverse-log plot...

> benchmarks; the hardware maxes out at 91-92 Mbps.

This is the point I wanted to clarify: Bogdan is talking about TCP
end-user payload data rates.  This _is_ maximum performance -- if
someone says "this NIC gets 97Mbps", they are measuring a different
number.

Every modern PCI NIC can get full wire rate performance.
The difference is how much system load the NIC imposes.
This is impacted by:
   Not requiring packet copies (the 3c905+ does not)
   Arbitrary Rx and Tx Packet alignment (very good)
   TCP/UDP/IP checksum calculations (the 905B/905C are excellent)
   Interrupt mitigation (very limited)
   PCI bursts per packet (mostly good)
   CPU bus operations per packet.
   The ability to tune FIFO thresholds before failures.

-- 
Donald Becker				[email protected]
Scyld Computing Corporation		http://www.scyld.com
410 Severn Ave. Suite 210		Scyld Beowulf cluster system
Annapolis MD 21403			410-990-9993

_______________________________________________
vortex mailing list
[email protected]
http://www.scyld.com/mailman/listinfo/vortex