Steve vainly tries to deflect deserved MI criticism:
> Stan Sieler ([log in to unmask]) wrote:
> : :fetch cicat.pub.sys (a file of 13456 sectors)
...
> : that's 50% slower! (Times on a 3000/968 with 128MB)
>
> : Ok, so you *DO NOT WANT TO EVER LEAVE A MI-BASED TOOL RUNNING ALL DAY*,
> : well...hardly ever :) (MI-based tools have their place, certainly!)
>
> Try your test on a Native Mode file type. Bet you can't get
> any appreciable difference.
The last time I looked, a file like:
ACCOUNT= SYS GROUP= PUB
FILENAME CODE ------------LOGICAL RECORD----------- ----SPACE----
SIZE TYP EOF LIMIT R/B SECTORS #X MX
CICAT 80B FA 42903 42903 16 13456 1 29
*IS* a "Native Mode file type".
Luckily for Steve, no wager was actually mentioned. :(
(I could have used that new memory board for our 3000/968)
As usual, I chose my examples carefully for reproducability (well, you
might not have "fetch"/KLONDIKE, but you can simulate the effects via
Debug's FVA command), and applicability (e.g., reading an ordinary file)
via an HP-provided & supported utility program that many shops use.
(Yes, it's often not appropriate to use FCOPY...but people still use it.)
I deliberately did not choose to read a CM KSAM file, an RIO file, or
a pre-5.0 MSG file ... which are disk files managed by CM code. (CIR files
pre 5.0 were managed by CM code too, but since that was MSG file code, I
assume (without checking) that they are now managed by NM code.)
Why? I knew that they tend to fare even worse when the MI is turned on.
I generally remember to point out that the MI impact seems to be
significantly worse for CM programs ... which includes FCOPY.
However, here's a completely NM test (as well as any program can claim
to be completely NM, given that various parts of MPE are in CM):
Program: REP.PUB.LPSTOOLS ... a high performance, NM copy program...
reads data via MR NOBUF reads.
Task: read CICAT.PUB.SYS (which has been frozen into memory)
and copy it into a new flat file (purging the prior one)
Major Intrinsics used:
FOPEN, FCLOSE (purge), FCLOSE (save), FREADDIR, FWRITEDIR
Timings:
CPU Elapsed (in milliseconds)
no MI 493 527
494 547
MI 522 567
523 557
MI slowdown: 5.8% 5.7% (using best times from each case)
(I ran the tests 16 times for each case, and reported the best two times
for each. The no-MI tests varied very little. The MI tests varied more.)
Please note:
During this period the MI-enabler (GLANCEXL) was *not* doing any work.
In other words, if you are actively collecting data and
logging it, that will incur a performance penalty above and beyond the
simple "is MI on" overhead ... which we've now seen ranges from an
easily demonstratable, non-controvertible 5.7% up to 50%. BTW, I've seen
worse percentages in the past (up to 70% overhead for extremely CM intensive
applications), but I quote them reluctantly as I don't think they are
representative of the average MI impact.
--
Stan Sieler [log in to unmask]
http://www.allegro.com/sieler.html
|