CPU profiling extension

[kolesnikovae]

In the world of high-performance systems, ensuring that CPU resources are allocated efficiently is crucial for achieving optimal performance.

However, mismatched CPU allocations, such as discrepancies in CPU capabilities between instances, cross-NUMA scheduling, or excessive core-sharing, can significantly degrade system performance and lead to complex, hard-to-diagnose issues.

To help users identify and mitigate these problems, I propose adding a CPU Allocation Analysis feature to Pyroscope.

The implementation mostly relies on the attribution we can identify in profilers. The basic attributes we may need to collect include:

• CPU model and stepping, ISA architecture
• Instance type (in cloud environments)
• Whether cross-NUMA scheduling occurs

Even attaching these basic attributes as profile labels can greatly help in investigating issues.

More advanced scenarios involve PMU analysis; for example, we could try to attribute cache misses to specific factors such as patterns in neighborship or hyper-threading contention.

A UI extension may follow to enhance usability.