Very strange measurements

[Eddie Kohler]

These are weird worrisome numbers!

> * however, I took click 1.0.6 (tested also click 1.1 and later, but
> these gave the same results)

There's a grammar ambiguity here. Was Click-1.0.6 the last release that
behaved well? Or is current CVS the first release that behaved badly? Which
other releases did you try?

I'm assuming that you meant Click-1.0.6 was the last release that behaved
well:

> Does anyone has an explanation for this behaviour (and what was the big
> chance between click 1.0.6 and later versions ?) or do we have to burn
> our GHz processors :-) ?

---

> it sounds very much like a scheduling thing.
> long latency due to some weird scheduling issue.

Benjie -- What scheduling issue in our code would differ between a
K6-500MHz and a K6-1GHz?

Brecht, the polling interface was significantly changed between 1.0.6 and
1.1b1. Benjie rewrote a lot of stuff.

My first guess: Are there any weird prefetching instructions remaining in
current Click when ENABLE_INTEL_CPU is not defined? Maybe K6-1GHz reacts
badly to these.

Brecht, more information, if you can? Can you run the PollDevice with its
profiling on (CLICK_DEVICE_STATS #defined), and tell us the results
(returned by "calls" handler)? Is PollDevice slower, or the device driver?

love,
ed