Skip to content

Latest commit

 

History

History
106 lines (81 loc) · 3.92 KB

README.md

File metadata and controls

106 lines (81 loc) · 3.92 KB

PMU Utils

Requirements

Currently, only Linux is supported. The library does not rely on any other library. It uses only standard C functionality.

Features

  • Support for x86_64 chips (Intel ia64 and AMD amd64) and ARM (aarch64) chips.
  • Support for reading and writing to the MSR register of the Performance Monitoring Unit (PMU).
  • Support for reading and writing the Performance Monitoring Counter (PMC) value.
  • Support for reading the Performance Monitoring Event (PME).
    • For x86_64 ia64 chips, the PME is getting from the intel/perfmon
    • For x86_64 amd64 and aarch64 chips, the PME is getting from the Linux Perf tool.

Read the following sections for more details.

Linux Usage

Preparation:

# Building the kernel module requires the kernel headers of the kernel. On Ubuntu or Debian, they can be installed by running
sudo apt install linux-headers-$(uname -r)

# The kernel module can be built by running
make

# The resulting kernel module can be loaded using
sudo insmod module/pmu_utils.ko

# Additionally step for x86_64 chips machines
sudo modprobe msr   # load msr module

# build the PMU Event Header for x86-64 chips (for `ia64` and `amd64` architectures), or for ARM chips
cd scripts
python3 pmu_intel_utils.py # for x86-64 ia64 chips
python3 pmu_intel_utils.py --hybridcore # if chips have hybrid core
python3 pmu_amd_utils.py   # for x86-64 amd64 chips
python3 pmu_arm_utils.py   # for ARM chips
# then you could see the PMU header: `ia64_pmu_event.h`, `amd64_pmu_event.h`, or `arm_pmu_event.h`
# just include the header in your c or c++ project

Using the library:
Refer to the tests directory for more details about how to use the library. All the functions are defined in libpmu/pmu.h.

#include <pmu_utils.h>

For x86_64 chips

uint64_t start_pmc = 0, end_pmc = 0;
uint64_t hexcode = 0x410000, core = 0, pmc_id = 0;
int msr_fd = pmu_open_msr(core); // open msr
pmu_set_msr_event(msr_fd, hexcode, core); // set event
pmu_record_start(msr_fd); // start record
pmu_set_pmc(msr_fd, pmc_id, 0); // set 0 for PMC[pmc_id]

start_pmc = pmu_get_rdpmc(pmc_id); // read PMC[pmc_id]
/*
 code that you want to measure
*/
end_pmc = pmu_get_rdpmc(pmc_id); // read PMC[pmc_id]

For aarch64 chips

uint64_t pmc_id0 = 0, pmc_id1 = 1;
pmu_set_event(0x0, NULL, evet_code[0], pmc_id0);
pmu_set_event(0x0, NULL, evet_code[0], pmc_id1);
memset(mem, 0, 4096);

pmu_values[0] = pmu_get_rdpmc(0);
pmu_values[1] = pmu_get_rdpmc(1);
/*
code that you want to measure
*/
pmu_values[0] = pmu_get_rdpmc(0) - pmu_values[0];
pmu_values[1] = pmu_get_rdpmc(1) - pmu_values[1];

Test and Example

Seen in tests directory.

cd tests && mkdir -p build && cd build && cmake .. && make tests

Script Utils

Change the directory to scripts and run *utils.py for your architecture.

References