Very strange measurements

[Eddie Kohler]

Hi Brecht,

> there's a word ambiguity here (:-) ) what do you mean by 'device driver'
> a simple ToDevice ?

hee!

There are four device-related things going on here:

      1. device driver (= Tulip) polling read
      2. PollDevice
      3. ToDevice
      4. device driver write

I want to know which have gone up unusually high, b/c it will limit our
search for problems.

It looks to me like the problem is in 1., the device polling read. We
normally don't see numbers this out of whack:

> 2210642777 cycles poll
> 41250394 cycles refill
> 159186084 cycles allocskb
> 1376208299 cycles pushing

Usually we see 'refill' taking 59 ns/packet, and 'poll' taking 497
ns/packet, ratio 'poll/refill' = 8.4. You are seeing ratio 'poll/refill' of
either 53.6 (!!) or 11 (I forget whether 'allocskb' was inclued in the
'refill' number). Either is high.

Maybe you could take these numbers for at least two loads, heavy (=low
latency result) and light (=hi)? And any ToDevice statistics? If it's not a
big deal?

Also, what is the router configuration?

lov,e
ed