TheRock/rocm-systems/projects/rocprofiler-systems/examples/parallel-overhead/README.md

Parallel Overhead

Overview

This example measures the overhead of parallel synchronization by comparing two approaches to shared state access: mutex-based locking and lock-free atomic operations. Multiple threads compute Fibonacci numbers with randomized inputs and accumulate results through either a mutex-protected shared variable or an atomic counter. By comparing the two builds, you can quantify the profiling overhead introduced by different synchronization primitives. Thread synchronization uses a pthread_barrier_t to ensure all threads start simultaneously.

Source Files

• parallel-overhead.cpp - Defines fib() (recursive Fibonacci, noinline), run() (per-thread worker with random input variance), and barrier-synchronized thread management. The USE_LOCKS compile-time macro switches between mutex and atomic modes.

Prerequisites

• CMake 3.25+
• C++ compiler with pthreads support

Building

Standalone build:

cmake -B <build_dir> -S <project_root>/examples/parallel-overhead
cmake --build <build_dir>

As part of the examples suite:

cmake -B <build_dir> -S <project_root>/examples/
cmake --build <build_dir> --target parallel-overhead parallel-overhead-locks

Target	Description
parallel-overhead	Lock-free atomic accumulation
parallel-overhead-locks	Mutex-based accumulation (USE_LOCKS=1)

Running

# Default: fib(10), min(16, ncpus) threads, 50000 iterations
./parallel-overhead

# Custom: fib(12), 8 threads, 20000 iterations
./parallel-overhead 12 8 20000

# Mutex variant
./parallel-overhead-locks 12 8 20000

Arguments:

Position	Description	Default
1	Fibonacci number to compute	10
2	Number of threads	min(16, hardware concurrency)
3	Number of iterations per thread	50000

Profiling with rocprofiler-systems

Compare overhead between variants:

# Atomic version
rocprof-sys-run -- ./parallel-overhead 10 8 10000

# Mutex version
rocprof-sys-run -- ./parallel-overhead-locks 10 8 10000

Recommended Configuration

Variable	Value	Purpose
ROCPROFSYS_TRACE	true	Generate Perfetto trace for timeline analysis
ROCPROFSYS_PROFILE	true	Generate call-stack profile
ROCPROFSYS_USE_SAMPLING	ON	Enable statistical sampling
ROCPROFSYS_SAMPLING_FREQ	50	Sampling frequency (interrupts/sec)