2 The Initial Run

In general, if you don't know much about the performance of the application you are studying, it is a good idea to perform an initial “profiling run” on a moderate number of nodes using representative input data. This is exactly what we are going to do with our IS benchmark.

Since we don't really have an idea of how time is being spent in the IS benchmark, it would be really great if PPW could provide us with a breakdown of how much time was spent executing different functions in the IS source code. To get this information, we make use of the --inst-functions flag. Go to the NPB2.4 directory that you set up in the previous section and type

     $ make UPCC="ppwupcc --inst-functions" IS NP=32 CLASS=B

or edit the make.def file as before. After a short delay, you should have a shiny new is.B.32 executable in the IS directory.

We can now run the is.B.32 executable directly as we normally would (srun/yod/upcrun/etc). If you do this, you should get a new data file in the directory where you ran the IS benchmark under a name like ppw-2234.par. This works OK, but it is more useful if we can tell PPW which filename to use. To do this, prefix your normal run command with the ppwrun command, such as

     $ ppwrun --output=is.B.par srun -N 32 is.B.32

Typing ppwrun with no options will show a short help screen that gives all the different options available for controlling PPW's measurement code at runtime. If you make use of the --output option, remember to give your data files an extension of .par to make them easy to find in the GUI later on.

Note: ppwrun works its magic by setting a bunch of environment variables, and relies on the job spawner to propagate these new environment variables when launching the job. If it seems like ppwrun is ignoring your options, then your job spawner probably doesn't handle propagating environment variables. If so, see the documentation for ppwrun in the user manual for some workarounds. Luckily, most job spawners do a good job of environment propagation.

Now that we've figured out how to run applications with PPW, go ahead and run the is.B.32 executable using ppwrun. Then, take the new data file created after running the benchmark and transfer it to your local workstation using scp or something similar. Start up the PPW frontend GUI, and open up the newly-created data file.

If you don't have an easy way of transferring data files from your parallel machine to your local workstation, then you may use the ppwprof and/or ppwprof.pl to view text-only performance information from the PAR data file. These commands provide only a subset of the functionality available from the PPW GUI, but are useful for quickly browsing performance data without having to transfer files around. The rest of this tutorial assumes you are using the PPW GUI to browse performance data.