Pmcount for Linux on Power Architecture

A hardware performance counter tool for the IBM POWER4, POWER4+, POWER5, POWER5+, and PowerPC 970 processors.

Date Posted: August 24, 2006

Overview Requirements Download FAQs Forum Reviews 1. Is the API POWER-specific? 2. Why do I need pmcount instead of OProfile? 3. What events are available? 4. Will this tool run on AIX? 5. What Linux distributions does the tool run on? 6. I get the error ERROR opening the EVS file .... What is wrong? 7. I get the error ERROR opening module loaded file. What is wrong? 8. The tool reports only counts for the even-numbered processors. Why? 9. Where can I get more information on Power Performance Counter Tool? 10. Why do I need the virtual 64-bit counter module? 1. Is the API POWER-specific? The current user/kernel API for setting the counters is a POWER4-, POWER5-, PowerPC 970-specific interface. We are working with the community to get a performance counter API that is architecture-independent and that is accepted into the Linux kernel. Once that happens, pmcount will be modified to use the new API. The plan is for the new API to provide the functionality that is currently provided by the kernel modules. After the new API is in place, the kernel modules will no longer be necessary. Back to top 2. Why do I need pmcount instead of OProfile? Oprofile does statistical sampling to collect addresses for a profile. It collects across the entire machine. Pmcount reports the actual performance counts. It can track a particular workload or "pid," or it can collect data for the entire machine. Although it is possible to infer approximately what the actual hardware counts were for a given OProfile run, the accuracy is limited by the number of events per sample selected. In addition, the overhead of collecting performance counter data with pmcount is much lower than with Oprofile. Back to top 3. What events are available? Use the command ./pmcount -L -1 in order to list the available events for the machine. You can use the --power option with the -L option to see the events for a different architecture. See the help guide for more information on the --power option. Back to top 4. Will this tool run on AIX? No; this tool was written to match the specification of the pmcount for AIX tool. However, it does not run on AIX. Back to top 5. What Linux distributions does the tool run on? The tool has been tested only on Red Hat and the SuSE distributions. Back to top 6. I get the error ERROR opening the EVS file .... What is wrong? The tool is not able to find the event file for the processor. Either change to the directory with the event and group files (these are the files with the evs and gps suffixes), or use the --dir option to specify the path to the directory where the files are located. Back to top 7. I get the error ERROR opening module loaded file. What is wrong? You must be root to run the tool. The tool was unable to find the kprobe modules. Be sure to use the --dir option if you are not running in the directory in which the tool was installed. Back to top 8. The tool reports only counts for the even-numbered processors. Why? The machine has the SMT state disabled. The SMT state can be enabled by specifying smt-enabled=on when booting the operating system. Back to top 9. Where can I get more information on Power Performance Counter Tool? A brief help guide is provided with the tool via the -h command line option. For example: ./pmcount -h. Also included is a user manual in PDF and in Postscript. See the installation directory for the user manual. Back to top 10. Why do I need the virtual 64-bit counter module? The hardware counters are only 32 bits. They will overflow in just a few seconds on a 1.6-GHz machine. The 64-bit counter module accumulates the counts from the 32-bit hardware counters so that no counts are lost due to overflowing of the 32-bit counters. Back to top