[E1000-devel] RxIntDelay and Packet Drop...

[Michael <mi...@tr...>]

hi all,

sorry for the question, but i somehow do not understand the RxIntDelay
parameter.

I have a Highend Machine whith several 82571EB GB Controllers.
This machine is a firewall, and a cat of /proc/net/dev shows that it
droppes a lot of packets.

The RxDescriptors and TxDescriptors are already set to = 4096, and
according to the vendor the RxIntDelay = 128
The machine recieves on each interface during peaktime approx 70K
Packets/Second, with higher peaks, and has
lots of idle cpu power..

My Support now insists that i modify the RxIntDelay from 128 to 1024 to
decrease the package drops..
( They claim that this is absolutely necessary )

But the Documentation from intel about this feature (
http://www.intel.com/support/network/sb/cs-009209.htm ) states:
"If the system is reporting dropped receives, this value may be set too
HIGH, causing the driver to run out of available receive descriptors."
( As the Machine is about 70% Idle, i do not car too much about cpu usage.. )

As the firewall is highly productive, i am not allowed to many downtimes,
so i am rather cautious about this. But as far
as i understand this issue the problem with higher RxIntDelay delay would be:

70K Packets/Second equals approx 7K Packets/Milliseconds, which means if i
add a 1 Millisecond delay i would
need 7K RxDescriptors (+ their buffers) to store the packets till the
Interrrupt kicks in and the cpu takes care
of the recieved packets.


Thanks in advance,
Michael