ndt-users list created

[Peter Van Epp <>]

On Tue, Jul 26, 2005 at 04:27:22PM -0400, Richard Carlson wrote:
> Hi Peter;
> 
> Thanks for the feedback.  I've used copper GigE cards before and never saw 
> this error, but that doesn't really tell you anything.  My guess is the NIC 
> is doing some pacing or packet bunching (maybe interrupt moderation).
>
<snip>

        He got it in one :-). Turns out this is an effect of the interrupt 
limiting timer being set to 8000 in the Intel E1000 driver. Setting that to
0 (which disables it at the cost of signifigantly increased CPU useage) fixes
both ndt and the Apparanet sequencer which was seeing odd UDP rate limiting.
        My older SysKonnect fibre cards (which also cure the problem) are too
old to do interrupt limiting so they don't exhibit the problem.
        Since running tcpdump on the same interface also seems to fix things
we think that turning promiscuous mode on with ifconfig may also work (which
wouldn't require unloading and reloading the E1000 driver which changing the
throttle setting does). The data sheet for the chip says that the throttle 
counter will delay the interrupt until it expires which would have the effect 
of presenting multiple packets with only a PCI transfer delay time between 
them (as the kernel timestamps the packet as it gets them from the driver) 
which makes it look like a 10 gig link because of bogusly short arrival 
times..
        As well we have discovered if you have a multihomed machine and start
ndt with -i eth0 -i eth1 it will happily run tests on which ever interface 
that
you point the test at (although this has obvious scheduling problems in the 
general case and may need a locking / queuing mechanism to serialize tests
in a less controlled environment).

Peter Van Epp / Operations and Technical Support 
Simon Fraser University, Burnaby, B.C. Canada