4. Profile Your Application

The following sections describe methods of profiling DPDK applications on different architectures.

4.1. Performance counter based profiling

Modern CPU architectures are equipped with Performance Monitoring Units (PMUs), which provide programmable counters to monitor specific hardware events, such as cache hits, instruction counts, and branch predictions.

Tools like Linux perf utilize PMUs to gather performance data. However, in scenarios where CPU cores are isolated, running dedicated tasks and performance of some specific regions of code must be analyzed, extra overhead may be undesirable. In such cases, applications can directly access PMU data using the rte_pmu_read() function.

4.1.1. Access requirements

Only Linux is supported.

Userspace applications may be restricted, due to various reasons, from accessing PMU internals. To enable access, /proc/sys/kernel/perf_event_paranoid should be set to 2 and application should have CAP_PERFMON capability assigned.

For comprehensive information on security implications and configuration, refer to Linux documentation.

4.1.2. Limitations

Current implementation imposes certain limitations:

  • Only EAL lcores are supported.

  • EAL lcores must not share a CPU.

  • Each EAL lcore measures the same group of events.

4.2. Profiling on x86

Intel processors provide performance counters to monitor events. Some tools provided by Intel, such as Intel® VTune™ Amplifier, can be used to profile and benchmark an application. See the VTune Performance Analyzer Essentials publication from Intel Press for more information.

For a DPDK application, this can be done in a Linux* application environment only.

The main situations that should be monitored through event counters are:

  • Cache misses

  • Branch mis-predicts

  • DTLB misses

  • Long latency instructions and exceptions

Refer to the Intel Performance Analysis Guide for details about application profiling.

4.2.1. Profiling with VTune

To allow VTune attaching to the DPDK application, reconfigure a DPDK build folder by passing -Dc_args=-DRTE_ETHDEV_PROFILE_WITH_VTUNE meson option and recompile the DPDK:

meson setup build
meson configure build -Dc_args=-DRTE_ETHDEV_PROFILE_WITH_VTUNE
ninja -C build

4.3. Profiling on ARM64

4.3.1. Using Linux perf

The ARM64 architecture provide performance counters to monitor events. The Linux perf tool can be used to profile and benchmark an application. In addition to the standard events, perf can be used to profile arm64 specific PMU (Performance Monitor Unit) events through raw events (-e -rXX).

For more details refer to the ARM64 specific PMU events enumeration.

4.3.2. Low-resolution generic counter

The default cntvct_el0 based rte_rdtsc() provides a portable means to get a wall clock counter in user space. Typically it runs at a lower clock frequency than the CPU clock frequency. Cycles counted using this method should be scaled to CPU clock frequency.

4.3.3. High-resolution cycle counter

The alternative method to enable rte_rdtsc() for a high resolution wall clock counter is through the ARMv8 PMU subsystem. The PMU cycle counter runs at CPU frequency. However, access to the PMU cycle counter from user space is not enabled by default in the arm64 linux kernel. It is possible to enable cycle counter for user space access by configuring the PMU from the privileged mode (kernel space).

By default the rte_rdtsc() implementation uses a portable cntvct_el0 scheme.

The example below shows the steps to configure the PMU based cycle counter on an ARMv8 machine.

git clone https://github.com/jerinjacobk/armv8_pmu_cycle_counter_el0
cd armv8_pmu_cycle_counter_el0
make
sudo insmod pmu_el0_cycle_counter.ko

Please refer to Compiling the DPDK Target from Source for generic details on compiling DPDK with meson.

In order to enable PMU based rte_rdtsc(), user needs to configure the build with -Dc_args='-DRTE_ARM_EAL_RDTSC_USE_PMU'.

Example:

meson setup --cross config/arm/arm64_armv8_linux_gcc -Dc_args='-DRTE_ARM_EAL_RDTSC_USE_PMU' build

Warning

The PMU based scheme is useful for high accuracy performance profiling with rte_rdtsc(). However, this method can not be used in conjunction with Linux userspace profiling tools like perf as this scheme alters the PMU registers state.