Ok, I have done some more digging and I have some more (limited)
information on this topic. See below...

> On Behalf Of Steve Cooper

> Scott McClellan <[log in to unmask]> wrote:
> >[Stan's demo script deleted...]
> >
> >For the sake of moving this debate in a more beneficial direction...
> >
> >Let's assume that the MI is slow (I really don't care what number you
> >want to assign to the overhead). [NOTE: this is an assumption not
> >an admission. I haven't measured it. For the record, Stan's test
> >looks credible enough to make one suspicious. I suspect the 44% is
> >a pseudo-worst case, however.]
> >
> >What is the business case for HP to spend R&D time and money enhancing
> >MI performance?

> I don't think there is one.  I think Stan (and I) would just like
> to see the false information that Glance's overhead is limited to the few
> percent that it itself sees, stop being repeated.
First of all I think you have expressed the key point in the middle of
your paragraph above. The MI overhead that some folks "quote" is the
overhead reported by Glance, which is not the "full story" as you and
Stan have pointed out. It is not that Glance is wrong, it's just that
all MI overhead does not show up as MI overhead. This information is more
missleading than is is out-and-out false. (Maybe I have been watching
the impeahment hearings too much! :-)

This entire mystery prompted me to keep digging, and in this case,
I kind of like what I found. It seems that HP has been listening a little
bit more than you and Stan want to believe (albeit their could well have
been more delay than you would have liked). I have it from reliable
sources that some significant improvements have been made, reducing
the MI overhead. As it is not my project, I would prefer to leave it
to the folks involved to post (or withold the details), but the data I
saw look much better. The bottom line it seems to me is:

a) the message was heard in the division (ultimately where it counts)
and some improvements were staffed, etc...
b) the was a business case, namely the performance implications are
very important to CSY's current business objectives...


> I have been paid more than once now to improve a system's performance,
> and the suggestion I make that has the highest payback, is to stop the
> continuous running of Glance or any other  product that leaves the MI
> enabled.  An occasional "peek" is fine; running it  when you are
> diagnosing a problem is fine.  But running it all of the time on
> a terminal in the computer room that hardly anyone ever looks at is a
> gigantic waste of system resources.
Well then I guess you owe us a cut on you consulting fees! ;-) All kidding
aside this seems like sound advice. I would not run Glance or any other
performance tool **ALL THE TIME** on my production system, especially
if peak performance were critical to me. One of you will probably say,
"...you can run XXX all the time 'cause it doesn't use the MI...". Well
my response is "maybe!". Basically, I don't have enough data to cause
me to believe that is a good idea.


> (Incidentally, I think the 44% is high, too, but I have seen
> numbers that high on real machines running alot of CPU-intensive CM
> processes.  Numbers is the 20-25% range are more typical, but that is
still
> throwing away  one fifth of your system.  I obtained these numbers by
> looking at total  metrics for a day (transactions, CPU seconds, etc.) with
> and without the MI enabled.)
Some overhead is inevitable, with any method of measurment (IMO).
An average overhead of 20-25% is un-necessarily high (IMO).
HP has made some improvments...I will leave for others to quantify,
give release details, or do whatever they feel is appropriate.