TheRock/rocm-systems/projects/rocprofiler-systems/examples/thread-limit

Thread Limit

Overview

This example stress-tests the threading subsystem by creating a large number of threads (up to 8000 by default) that each perform recursive Fibonacci computations. Threads are launched in batches matching the hardware concurrency level, with each batch completing before the next starts. The program reports average execution time per thread and total throughput, making it useful for profiling thread creation/destruction overhead, scheduler behavior, and profiler scalability under high thread counts.

Source Files

• thread-limit.cpp - Implements batched thread creation using std::thread, each computing fib(n) recursively. Uses std::mutex to accumulate per-thread timing into a global total. MAX_THREADS is configurable at compile time (default 4000).

Prerequisites

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

Building

Standalone build:

cmake -B <build_dir> -S <project_root>/examples/thread-limit
cmake --build <build_dir>

As part of the examples suite:

cmake -B <build_dir> -S <project_root>/examples/
cmake --build <build_dir> --target thread-limit

Running

# Default: fib(35), concurrency=ncpus, 8000 total threads
./thread-limit

# Custom: fib(30), 8 threads per batch, 2000 total
./thread-limit 30 8 2000

Arguments:

Position	Description	Default
1	Fibonacci number to compute	35
2	Concurrency (threads per batch)	hardware_concurrency()
3	Total number of threads to launch	2 * MAX_THREADS (8000)

Profiling with rocprofiler-systems

rocprof-sys-run -- ./thread-limit 30 4 500

Recommended Configuration

Variable	Value	Purpose
ROCPROFSYS_TRACE	true	Generate Perfetto trace showing thread lifecycle
ROCPROFSYS_PROFILE	true	Generate call-stack profile
ROCPROFSYS_USE_SAMPLING	ON	Enable statistical sampling of thread stacks
ROCPROFSYS_SAMPLING_FREQ	50	Sampling frequency (interrupts/sec)