gige benchmark performance

Mark Hartner hartner at cs.utah.edu
Thu Mar 14 19:40:57 PST 2002


> well, the driver's author just posted an update that looks like 
> a major improvement.  he also says he's seeing 810 Mbps between 
> two athlons using 1.5KB MTU's.  which leads me to my second comment:
> that I've heard people say bad things about the i860's PCI throughput.

Hmm...

We didn't test the PCI throughput, but we did run some tests on the AGP
bandwidth. In particular, we ran attempted to saturate the AGP bus to see
how it affected network performance and vs versa. If you are interested in
seeing those, they are at www.cs.utah.edu/~hartner/cluster/cluster.html

We found the 860 chipset did quite well. We also ran tests on an Athlon
system (Tyan S2460), but I was never able to get the graphics stable on
that system.


> I don't know how that would explain the apparent collapse shown
> in your graphs, though.  ie, why the graphs start to look strange at 
> around 10KB packets.  I suppose there might be some interaction between
> back-to-back packets, the HW or driver's attempts at irq mitigation,
> perhaps even driver memory allocation.

The really odd thing about the collapse is that it only happens on the
64bit PCI slots. I just ran the numbers on one of the 32bit PCI slots and
the droppoff does not happen, and I get slightly better performance
overall. I don't think it is an IRQ issue either becuase I assigned the
same IRQ to the 32bit slot as the 64bit slots. I also made sure there were
no IRQ sharing issues. I will post these numbers if anyone is interseted.

For the Intel Pro/1000T and Netgear GA620 there was only a slight 
performance difference between 32 and 64 bit PCI (the 64bit slot did
slightly better).

Mark




More information about the Beowulf mailing list