[openib-general] [patch] misc trivia nitpicks 04

[Grant Grundler]

On Sat, May 07, 2005 at 01:51:35AM +0200, Bernhard Fischer wrote:
> Other notes:
> 
> getopt 'i' fills in ib_port but checks port, it seems. Didn't look
> closely.

Yeah, that looks wrong to me too. I think it's just
trying to validate that ib_port is a reasonable number.

> iters is off by one (stops at 998 when asked for 1000).

Yes, but not like you think.  "iters" runs from 0-998 - ie 999 samples.
Is adding one to the -H output the right answer?
(ie run 1-999)

We asked for 1000 "exchanges".  The measured the time between them
results in one less sample. The usage says 
    -n, --iters=<iters>    number of exchanges (at least 2, default 1000)

I'm inclined to leave that alone.

...
> 994, 8.759292
> 995, 9.820456
> 996, 26.436613
> 997, 72.905357
> 998, 704572144.206260

998 is probably garbage.
I'll sort that out on the ia64 box.

Several things could be causing this besides program error:
- process got moved to another cpu between samples.
  taskset should take care of this.
- CPU clock rolled over and the conversion from unsigned long long
  to double happened before the difference was taken.

And apologies for not posting the -02 tarball...got distracted
by some other issues. Almost done...

hth,
grant