13. Performance Debug using AMDuProfSys

13.1. Overview

AMDuProfSys is a system debugging and analysis tool for AMD processors. It can be used to collect hardware events and evaluate the simple counter values or complex recipes using collected raw events. The performance metrics are based on the profile data collected using Core, L3, DF, and UMC PMCs (Performance Monitoring Counters) or HSMP messaging (via SMN interface).

This tool provides the overall performance details of the hardware blocks used in the system.

13.2. Supported Platforms

AMDuProfSys supports AMD EPYC™ 7002, 7003, and 9000 Series processors with the following variants:

• Family 17, model 0x30 – 0x3F

• Family 17, model 0x60 - 0x6F

• Family 19, model 0x0 - 0xF

• Family 19, model 0x1 - 0x1F

• Family 19, model 0x20 - 0x2F

• Family 19, model 0x60 – 0x6F

• Family 19, model 0x70 – 0x7F

• Family 19, model 0x90 - 0x9F

• Family 19, model 0xA0 – 0xAF

• Family 1A, model 0x0 – 0xF

• Family 1A, model 0x10 – 0x1F

• Family 1A, model 0x20 – 0x2F

• Family 1A, model 0x40 – 0x4F

• Family 1A, model 0x60 – 0x63

• Family 1A, model 0x70

13.3. Supported Hardware Counters

The supported hardware counters are listed here. - CORE PMC - DF PMC - L3 PMC - UMC PMC - HSMP (via SMN interface)

13.4. Supported Operating Systems

The supported operating systems are:

• Linux

• Windows

13.4.1. Prerequisites

13.4.1.1. Windows

To profile L3 and DF counters while Hyper-V is enabled, switch the system to system mode using the following command. After execution, reboot the system.

• To switch to system mode:

  bcdedit /set hypervisorperfmon system
  

• To switch back to default:

  bcdedit /deletevalue hypervisorperfmon
  

13.4.2. Set up

Execute the OS-specific installation steps provided in this section.

13.4.2.1. Linux

If using tar ball, install uProf driver manually.

To install the AMDuProfSys driver:

#  ./AMDPowerProfilerDriver.sh install

Alternatively, Linux Perf can also be used if the user space perf tool is installed in the system. Make sure that Perf tools support the required PMC event monitoring for the platform used. To use Linux Perf for collection instead of the uProf driver, include the --use-linux-perf flag in the command.

To install user space perf tool:

# apt-get install linux-tools-common linux-tools-generic linux-tools-`uname -r`

Disable NMI watchdog; this requires root privileges:

# echo 0 > /proc/sys/kernel/nmi_watchdog

Set Perf parameter to -1 if system-wide profile data or DF and L3 metrics must be collected:

# sh -c 'echo -1 >/proc/sys/kernel/perf_event_paranoid'

To collect using Linux Perf, check if the amd_uncore module is loaded using the following command:

$ lsmod | grep amd_uncore

If not, run the following command:

# modprobe amd_uncore

After successful installation, you can access AMDuProfSys from <Installed Directory>/bin/AMDPerf/AMDuProfSys.

13.4.2.2. Windows

The Setup file installs all the necessary components required to run AMDuProfSys. Once the installation is complete, AMDuProfSys can be accessed at:

<Installed Directory>/bin/AMDPerf/AMDuProfSys.exe

13.5. Options

13.5.1. Synopsis

AMDuProfSys [-h] [--help-all] [--version] <COMMAND> <options> <WORKLOAD> <workload-specific-args>

where:

• <COMMAND> — To collect, generate report, or get help for this tool

• <options> — Command-specific options; detailed in the following tables

• <WORKLOAD> — Denotes a launch application to be profiled

• <workload-specific-args> — Denotes the list of arguments for the application to be launched using AMDuProfSys

13.5.2. Generic

The following table lists the generic options.

13.5.3. Collect Command

The Collect command is used to configure profiling and monitor raw events while the launched program is executing, the duration set expires or Ctrl+C key combination is hit.

The options listed in the following table can be preceded by a “collect” keyword for collection of profile data. To collect and generate metrics reports in a single step, the same options can be used with the collect keyword omitted. In this case, report command options are also supported.

The following table lists the collect command options.

13.5.4. Report Command

The report command is used to generate a profile report with computed metrics. The collect command run prior to report generation will have generated a profile session file with .ses extension and a raw counter data file for each type of metric-type profiled.

To generate the report, the path to the session file needs to be provided as input (with -i option) as shown in the following command options:

13.6. Configuration Option

13.6.1. Default configuration

Without any config option, a minimal set of metrics (based on core, l3 and df) are collected. This may help get a firsthand system analysis. The following are some of the metrics data that can be collected.

Example command: AMDuProfSys -o /tmp/default -d 30

This will generate the default directory containing default_core.csv, default_l3.csv, and default_df.csv.

13.6.2. Predefined Configurations

AMDuProfSys bundles a bunch of predefined config files for core, l3, df, umc, and hsmp. These configuration files are in json and yaml formats which includes a detailed list of metrics per the metrics type.

Example command: AMDuProfSys --config core,l3,df -a -d 5 -o /tmp/all

This will generate the all directory containing all_core.csv, all_l3.csv, and all_df.csv files.

13.6.3. Obfuscation Support

AMDuProfSys allows obfuscated events to be collected on user systems without needing to expose internal events properties. Users are allowed only to collect the profile data. For generating reports, users must send the collected profile data to AMD.

13.6.3.1. Feature Workflow

1. Collect obfuscated profile data: Users can collect raw config data using raw_config file provided with the AMDuProfSys package.

   Example command

   AMDuProfSys collect --config core --obfuscation -o temp -d 30
   

   This generates a session folder containing a session file (.ses) and a raw file.

2. Send data to AMD: Users must send the entire session folder (containing all files) to AMD.

3. Report generation: Profile reports can only be generated using the internal version of AMDuProfSys by AMD. The session file (.ses) is used as the input to generate the report.

   Example command

   AMDuProfSys report -i ./temp/temp.ses
   

13.6.4. Custom Metrics

Use this option to write your own custom metrics and profile only for those metrics. This option helps exclude all unwanted metrics from the options mentioned earlier to minimize the overhead of multiplexing. However, this option requires knowledge of PMC encoding and knowledge of metrics that you want to profile for. Here is an example of metrics along with an AMDuProfSys command to get the profile data.

TotalRetiredBranceNotTaken= total retired branch- retired branch taken
= PMCx0C2 – umask 0x0 - PMCx0C4 – umask 0x0
= 0x4300C2-0x4300C4
RetiredBranceNotTaken = (TotalRetiredBranceNotTaken * 1000) / Total retired instructions
= (TotalRetiredBranceNotTaken * 1000) / 0x4300C0

Example command

AMDuProfSys --metrics   core/TotalRetiredBranceNotTaken="0x4300C2-0x4300C4",core/RetiredBranceNotTaken="(TotalRetiredBranceNotTaken*1000)/0x4300C0"

Output

Examples

13.7. Limitations

The number of events that can be collected at a time is restricted by the number of hardware PMCs available on a platform. AMDuProfSys divides the list of events into groups, with each group consisting of a maximum of PMC-count number of events. These events groups are multiplexed during a run. Due to this, sampling may not provide 100% accurate data, especially if the profiling has run for very little time.

• L3 metrics are not available for profiling on the following platforms:

  • Family 17, model 0x60 – 0x6F

  • Family 19, model 0x20 – 0x2F

• DF metrics are not available for profiling on the following platforms:

  • Family 17, model 0x60 – 0x6F

  • Family 19, model 0x20 – 0x2F

  • Family 19, model 0x60 – 0x6F

  • Family 19, model 0x70 – 0x7F

  • Family 1A, model 0x60 – 0x63

• UMC metrics are not available for profiling on the following platforms:

  • Family 17, model 0x30 – 0x3F

  • Family 17, model 0x60 – 0x6F

  • Family 19, model 0x0 – 0xF

  • Family 19, model 0x20 – 0x2F

  • Family 19, model 0x60 – 0x6F

  • Family 19, model 0x70 – 0x7F

  • Family 19, model 0x90 – 0x9F

  • Family 1A, model 0x60 – 0x63

13.8. Virtualization Support