HP3000-L Archives

January 2001, Week 3

HP3000-L@RAVEN.UTC.EDU

Options: Use Monospaced Font
Show Text Part by Default
Show All Mail Headers

Message: [<< First] [< Prev] [Next >] [Last >>]
Topic: [<< First] [< Prev] [Next >] [Last >>]
Author: [<< First] [< Prev] [Next >] [Last >>]

Print Reply
Subject:
From:
Gavin Scott <[log in to unmask]>
Reply To:
Gavin Scott <[log in to unmask]>
Date:
Wed, 17 Jan 2001 09:20:11 -0800
Content-Type:
text/plain
Parts/Attachments:
text/plain (45 lines)
Patrickasks:
> I found this interesting. It sounds like a good argument against running
> performance measurement tools all the time. And I know Stan has crusaded on
> the list for some time about the evils of the Management Interface and the
> amount of overhead it uses/causes. But is there an alternative for
> collecting performance stats? AFAIK HP has basically denied there's a
> problem with the MI (or at least said, "We're not changing it."). So what
> can we poor lil' sys admins do who need to be able to report on system
> performance on a daily/weekly/monthly basis that doesn't itself add a whole
> lot of overhead?

The open question is still how much overhead the MI actually adds on average
over a whole system running a typical 3000 application mix.  We've seen
enough examples of individual things that the MI slows down that I've slowly
come to agree with Stan's opinion of the MI, and it would not surprise me if
the MI *could* add up to around a 10% slow down system wide.

The only way to tell would be to run some much larger and more complicated
"whole system" benchmarks, which nobody has done yet (that I know of).

Of course whatever the number is, it will vary from system to system based
on how your application is written.  It's also quite possible that the total
overhead may not be this large for most systems.

I wouldn't suggest that one should *never* use MI based tools, just that you
probably shouldn't run them all the time if you can avoid it.  Of course if
the overhead is significantly less than 10%, then the information that
Glance and other MI tools gives you might easily be worth it to you to spend
that performance in order to get that information.

As far as non-MI ways of getting performance information, it depends on what
you want to see.  There are things like run time and CPU usage that the
system always collects, and there is Stan's SHOT program (available from
Lund Performance) that provides a lot of Glance-like information without
using the MI, but there are probably some things that can only be obtained
from the MI.

For things like response time, it may be relatively easy to instrument some
of your application programs to simply use calls to the TIMER Intrinsic
around "transactions" to explicitly collect the information that you want
and dump it somewhere that you can look at it after the fact (or in real
time if you want to get clever).

G.

ATOM RSS1 RSS2