Cuda Kernel Interference Performance

This repository provides a standardized framework for CUDA kernel performance testing and execution management.

Core Components

KernelBase Class

Base class exposing the following key methods for standardized kernel testing workflows:

• get_param_list(): Retrieves configurable kernel parameters
• set_params_and_prepare_input(): Applies runtime parameters
• prepare_input(): Handles input data preparation
• launch_kernel(): Executes the CUDA kernel
• profile_kernel_us(): Measures kernel execution time (μs)
• get_best_performance_us(): Returns optimal performance metrics

Resource Management

Built-in interfaces for:

• CUDA stream management
• SM (Streaming Multiprocessor) allocation
• Enables performance testing under varying compute resource constraints.

Implementation Guide

To integrate a new operator:

• Inherit from KernelBase
• Implement all required methods
• Refer to the reference implementation: mla/kernel_mla_decode.py

Single kernel test

cd test
python3 mla_decode.py

Interference test

cd test/interference
# if meet errors about cuda sync, try CUDA_LAUNCH_BLOCKING=1
python3 ./interference_test.py