[Orca-dev] Orcallator.se changes and additions

[Dmitry Berezin]

I would like to suggest a couple of changes to the orcallator.se script as
well as make some additions to it.

 

1. "Fix" timestamp.

            Timestamps that orcallator.se uses for each data entry do not
perfectly lineup with the data gathering interval and could be off by a
couple of seconds. In case when the data represents a whole quantities (such
as number of http processes or number of database connections) this creates
erroneous results (30.998 database connections for instance). This problem
could be avoided by fixing the timestamp to the interval boundary:    

    now    = time(0);

    now = (now/interval)*interval;

I agree that "fixing" any data is not a very good idea, but since the change
in the value of each timestamp will be small, I do not think that this will
have any significant effect on data averaging.

 

2. Problem with compressing data files.

            There are two places in orcallator.se where it calls "compress"
program defined in the environment. Both times it is executed on the
background. If the header information changes and log file needs to be
switched, orcallator.se first closes that log file and calls "compress",
then goes to search for the new available log file name, compressing all
uncompressed files that it finds. The problem is that the log file that was
just closed is still being compressed on the background at this time, so
orcallator.se calls another "compress" for the same file. Obviously, one of
these calls fails. I suggest changing orcallator.se, so that when it closes
a log file, it compresses it on the foreground.

 

3. Problem with proc_next().

            It appears that first_proc() / next_proc() combination doesn't
always traverse through all of the processes on the system. Every 10th -20th
time it stops somewhere in the middle. I have ran some tests on Solaris 8
with SE 3.3.1 that consistently confirmed that behavior. Here are the two
loops that go through all processes - the second one seems to be consistent
(orcallator.se uses the first one). 

 

*******************************************************

  for (p=first_proc(); p.pr_pid != -1; p=next_proc()) {

  }

*******************************************************

  proc_class_t pp;

  while (pp.index$ != -1) {

    refresh$(pp);

  }

*******************************************************

 

These three are minor changes and I have already implemented and tested them
on my server. Blair, I can send the diffs for all/any of these if you would
like to merge them into the official release.

 

4. The next topic is Disk Statistics.

            The RAWDISK code appears to be the biggest problem for the new
installations of Orca, especially on the newer hardware. It also appears
that it DOES NOT bring any advantages, since it was written for a very
specific situation, and the original code handles everything well enough. I
would like to know if anybody is still running orcallator.se on the servers
where RAWDISK code is necessary, and if not, it would probably be wise to
remove it (or at least disable it by default). I am also going to write a
patch for tape (st) statistics collection that will not depend on RAWDISK
code being enabled and post it here after testing it.

 

5. Process count.

            Orcallator.se provides the ability for counting http and https
processes, but in a lot of cases there might be a need to count some other
types of processes (oracle database connections, for example). It would
probably be better to have a generic function that would count "any"
processes based on some initialization variables.

Blair, is this of any interest. If yes, I could work on this. However, see
the next topic

(I have also written a patch that counts Oracle database connections for all
Oracle databases, dynamically finding all instances; but now I am not sure
if this is of any value to anybody.)

 

6. Workload statistics.

            I have written a new class based on the workload class that now
comes with SE Toolkit. The new class is more "orca friendly" and allows for
easier initialization (IMHO). I have also written a patch for orcallator.se
to use this new class to collect workload statistics. Coincidentally, the
new class provides the ability to get process counts (based on all/any of
the following: program name, program arguments, user name, processor set
binding) as well, so it could be used to count the number of http processes,
oracle processes, etc.

Blair, I can send this for review. Also, I need some help with graphing the
data.

 

  -Dmitry.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: /pipermail/orca-dev/attachments/20040511/92f926c1/attachment.htm