Subject: Re: interrupt rate from a NIC
To: Michael Graff <email@example.com>
From: Jonathan Stone <jonathan@DSG.Stanford.EDU>
Date: 05/23/2003 14:07:36
Michael Graff writes, about livelock:
>It doesn't occur at 100 Mbps or even gigabit traffic. It happens when
>you send lots and lots of minimal packets. A DS3 can take down a 100
>Mbps connected Pentium-3 866 Mhz. I know this from experienced.
Otoh, for well-designed NIC hardware[*], it is a SMOP to tune the
interrupt rate as a function of CPU load: higher load, interrupt less
often (even if you're being hit by tinygrams).
Just another wrinkle on Jeff Mogul's solution to livelock: switch to
polling under heavy load.
[*] Really deep packet DMA buffers, say 512 or 1024, plus the ability
for software to tune interrupt deferral without a chip reset or
something equally gross.