PMU overview

The performance monitoring unit (PMU) is a hardware component within the CPU core that monitors how the processor runs code. When you enable the PMU in an A4X, C4A, C4, or M4 Compute Engine instance, you can access the performance counters in the PMU by using performance-monitoring software. This approach lets you optimize performance-sensitive workloads, such as high performance computing (HPC) or machine learning (ML), by helping you identify and address performance bottlenecks in your applications.

This document explains the behavior, billing, and limitations of the PMU. To learn how to enable the PMU in an A4X, C4A, C4, or M4 instance, see Enable the PMU in Compute Engine instances .

Understand the PMU

The PMU is composed of a set of hardware counters called performance monitoring counters (PMCs). These counters are model-specific registers that count each time a low-level processor event occurs within the CPU, such as a branch misprediction or cache miss. You can read and configure PMCs in the PMU by using performance-monitoring software such as Intel VTune Profiler .

After you enable the PMU in a compute instance, the PMU runs in the background, continuously monitoring performance events using PMCs. You can optionally configure thresholds for specific PMCs using your preferred performance-monitoring software. If a PMC exceeds its designated threshold, then the PMU notifies the software.

PMU events

By default, the PMU is disabled in compute instances. To enable it, specify the types of low-level CPU events to track by enabling one of the following PMU types:

• Architectural ( ARCHITECTURAL ): you can measure the following architectural performance events:
  • Branch instructions retired: the number of branch instructions retired. Use this event to measure your code's execution and identify potential performance bottlenecks.
  • Branch misses retired: the number of branch instructions that were mispredicted, causing the processor to stall and discard fetched instructions. If you see a high number for this event, then you can likely optimize the CPU performance.
  • Instructions retired: the number of instructions the CPU successfully processes. Use this event to measure the CPU's instruction throughput.
  • Top down slots: the number of available slots within a processor's pipeline that are used to simultaneously execute instructions. Use this event to understand how efficiently your code is using the processor's resources.
  • Unhalted core cycles: the number of core cycles when the thread isn't halted—for example, due to power management or interrupts. Use this event to evaluate the overall usage of the processor.
  • Unhalted reference cycles: the number of reference cycles when the core isn't halted—for example, when fetching data or instructions. The core is halted when it runs the HLT or MWAIT instructions . Reference cycles operate at a fixed frequency, providing a stable time reference even when the speed of the processor changes to preserve energy. Use this event to measure the time spent on a task and identify performance bottlenecks in your code.
• Standard ( STANDARD ): you can measure all events from the Architectural PMU type and any local events inside the CPU core, including level 2 (L2) cache events.
• Enhanced ( ENHANCED ): you can measure all events from the Standard PMU type and any local events outside the CPU core, including level 3 (L3) cache events.

Supported machine series and CPU platforms

You can enable the PMU on compute instances that use the following machine series and CPU platforms:

Supported PMU events by CPU platform

The following tables list the supported PMU events by CPU platform:

Pricing

There are no costs associated with using the PMU in a compute instance.

Limitations

You can enable the enhanced PMU type only in C4 or M4 instances that use one of the following machine types:

• Any C4 machine type with 144 or 288 vCPUs

• One of the following M4 machine types:

  • m4-megamem-112

  • m4-megamem-224

  • m4-ultramem-56

  • m4-ultramem-112

  • m4-ultramem-224

What's next

• Enable the PMU in Compute Engine instances

• Enable the PMU in Google Kubernetes Engine clusters