On Tue, 27 May 2008 15:02:15 PDT "ron minnich" <[EMAIL PROTECTED]> wrote:
> OK, this is a long shot, but i'm running out of ideas.
>
> Long, long ago, at a Usenix, I saw a talk by some adventurous
> australians (are there any other kind?). It was concerning some neat
> hardware designed for kernel monitoring.
>
> They had done a very neat hack. Basically, they modified the C
> compiler so that, on function entry and exit, the code would emit a
> 16-bit quantity to the parallel port. They had some simple hardware to
> grab the data.
>
> WIth this, they were able to get some nice kernel performance numbers,
> all for the (low at the time) cost of an outw to the parallel port.
>
> OK, I have done some searching and can't find this. IIRC it was
> pre-website usenix. I am going to UCB this week and may have time to
> hunt it down in the paper archives, but ... just wondering ... anyone
> else remember this?
Not quite the same but perhaps you are thinking of "Hardware
Profiling of Kernels" by Andrew McRae in Winter 1993 usenix.
>From his paper:
Three basic building blocks are used in the proling system
proposed; the first is a hardware device that is used to
record time and event data into a RAM block. The second is a
modified C compiler that allows event triggering code to be
inserted into key locations, and finally the last building
block is analysis software that is used to decode the
backtrace of events and relate it to the source code.
http://mcrae.homeunix.net/papers/final_usenix.pdf
He plugs his simple h/w in an Eprom socket and captures upto
16K events. Each event has a 16 bit tag and 24 bit us clock
value (so events can be upto 16 seconds apart). Makes more
sense as access to a parallel port would be much slower. For
today's processors, this would be far too slow.
