application Responce Measurement API's

[David Fosgate_1]

I have a number of the performance tools running now. PerfView, Glance appear to report down to the PID level as the lowest level in regards to processes running on the system. If the process is hogging the system, I can see files that are open a as well as other metrics. The server still doesn't report the application program running. To add this missing information HP suggests using the API's.

Has anyone used either the older Transaction Tracking API or ARM API's? If so , how well do they work? Do they impact the application performance? Are they best used only during development to benchmark? Do they add enough information to the metrics reported to MWA to better focus on the bottlenecks?

Any real life experiences would be appreciated.

Thanks

Dave

[David Fosgate_1]

Oh Ya, and what types of applications have you installed the API's. C, Perl, Power Builder, etc?

[Jim Welch · Accepted solution]

Actually glanceplus will report down to the thread level. I'm not sure what exactly you mean by the following statement:
"The server still doesn't report the application program running."

You can certainly configure an application defination in the measureware parm file and lump any process/thread metrics into the application metrics.

For background info on the ARM api you might want to go to http://www.cmg.org/regions/cmgarmw/

There are also a couple of docs in /opt/perf/paperdocs/arm/C
(ARM 2.0 api guide, and Tracking Your Transactions)

ARM is not supposed to use a lot of overhead, but it could depend on your application and you transaction rate. Also, if your app is java based you might want to wait for the 3.0 api. It shouldn't be too hard to conditionally compile the arm calls into the code for a couple of transactions and benchmark it with and without the ARM calls to see what effect it has on the app's performance.

[David Fosgate_1]

Jim et all,

As a newbie to the unix os, I'm noticing that one user PID can spawn many processes (threads) for an application. What I'm looking to do is find a way to group any and all processes initiated by a users job and have this metric reported in Glance, Measureware. It appears the way to do this is an application definition. Other than the paperdocs information, is there another place to get in-depth information on creating application definitions for performance tools.

[Jim Welch]

The MeasureWare applicaton defination tells MeasureWare to group the resource utilization for processes (including the threads created by those processes) into the application metrics. There is not a lot to the application defination itself and the syntax is covered in the user's manual (starts on page 22 in the current version). The biggest 'gotchas' are that the file parameter only uses the process name, not any parameters used when the process was involked, so if you wanted to assign oracle processes to different applications based on some command line parameter passed to the oracle process, you can't. You can only use the process name, the userid, groupid, and priority to select which processes are included in a particular application defination. If you could control what user starts the process in the oracle example, then you could sort them into different applications. Also, the parm file is parsed from beginning to end and the first application defination that a process fits into is the one that the process's resource usage with be allocated to, so if you add a defination to the end of the parm file any processes that you intended to go into that application may end up in one of the ones above it, like other_user_root. Once you modify the parm file, you can run glance or gpm to a quick idea of how well the new defination is working. To get MeasureWare to recognize the changes to the parm file, you would need to restart MeasureWare (mwa restart).
If you have any other questions we can answer in a paragraph or two, you can post them here.